Physical AI Brief

Recent robot foundation models operate with single-step or short-history visuomotor context. We introduce Test-Time-Training Robot Policies (RoboTTT), a robot model and training recipe that scale visuomotor context to 8K timesteps, three orders of magnitude beyond state-of-the-art policies, without growing inference latency. At this context length, we unlock new robot capabilities: one-shot in-context imitation from human video demonstrations, on-the-fly policy improvement, robustness to perturbations, and stronger performance on multi-stage, long-horizon tasks. We also observe, for the first time, steady gains in closed-loop performance as pretraining context length scales. At its core, RoboTTT integrates Test-Time Training into robot foundation models such as Vision-Language-Action policies, yielding a sequence model whose recurrent state consists of fast weights, parameters updated by gradient descent during both training and inference, compressing histories into weight space and retrieving contextual information for long-context conditioning. To scale training context length, the recipe combines sequence action forcing with truncated backpropagation through time. On challenging real-robot manipulation tasks, RoboTTT improves overall performance by 87% over the single-step context baseline and fully completes a five-minute, ten-stage assembly task, which no baseline ever does. RoboTTT trained with 8K-timestep context outperforms the same model pretrained with 1K timesteps by 62%, suggesting context length as a new scaling axis for robot foundation models. Videos are available at https://research.nvidia.com/labs/gear/robottt/

Editing the figures in a research paper is a routine and time-consuming part of everyday research practice: authors relabel components, rearrange panels, and restyle visuals as they revise their manuscripts. Automating this editing workflow under a natural-language instruction, however, is challenging, because a scientific figure is a dense infographic in which heterogeneous visual elements such as schematics, plots, photos, captions, and arrows are composed under a tight visual grammar to advance a specific argument. To address this, we present SciDiagramEdit, a benchmark and skill-evolution framework that learns from natural paper revisions and operates on the figure's editable vector source, where users can inspect and co-edit individual primitives alongside the agent. Our benchmark mines before/after figure pairs from arXiv version histories, each grounded in the authors' own revision intent. To accommodate the diversity of editing instructions, we adopt agentic learning via skill evolution: an agentic proposer continually refines the agent's skill specification from execution traces over multiple epochs. The resulting skill progressively lifts edit accuracy on a held-out validation set, providing evidence that natural paper revisions are an effective training signal for instruction-driven figure editing.

Online novel view synthesis from multi-view streaming videos faces a fundamental trade-off: maintaining a persistent, long-horizon memory to reconstruct temporarily occluded regions while operating under strict real-time constraints. While Test-Time Training (TTT) offers a powerful memory mechanism, standard models mandate gradient-based memory updates at every frame to adapt to the changing motion in dynamic scenes. The computational cost of heavy memory updates precludes real-time application and can lead to instability over long contexts. Given that memory updates are more demanding than memory application and video content is largely redundant, we propose to decouple the frequencies of these two processes. Our approach performs periodic memory updates while applying the memory on a per-frame basis, using cross-view attention to manage deformations between the prior memory state and the current frame. To lock in the historical context, we introduce two critical mechanisms: an auxiliary Memory Loss that forces persistent internalization of the scene, and a Memory Caching strategy that regularizes active weights against catastrophic drift. Our method demonstrates real-time, state-of-the-art performance on scenes with dynamic human motion as well as minute-scale online memorization.

Security-agent evaluations commonly measure peak offensive capability under generous inference budgets, emphasizing vulnerability discovery, exploit development, penetration testing, and CTF completion. Such measurements are useful but incomplete: in operational security, every reasoning step, tool call, telemetry query, and enrichment request consumes budget. We evaluate language-model security agents through this cost-success lens on offensive Cybench challenges and defensive Splunk BOTS v1 investigation challenges. Instead of reporting only best-case success, we compare models at fixed cost levels and decompose performance by inference spend and tool spend. Our results show distinct scalingregimes for red- and blue-team tasks. Offensive CTF performance improves with additional test-time compute, and scaled open-weight models can approach frontier proprietary systems while remaining cost-competitive. Defensive SOC investigation does not scale in the same way: success depends more heavily on disciplined tool use, telemetry navigation, and selective enrichment than on raw reasoning budget alone. We argue that security-agent benchmarks should measure economic efficiency and operational fit alongside task success. Cost-aware, SOC-native evaluations provide a clearer picture of which models are practically useful today and where defensive agents still need to improve. We present an interactive website with our results https://evals.frontier.security.

Recent advances in Tool-Integrated Large Language Models have made web search a core capability of information-seeking agents. However, as interaction histories grow, agents increasingly struggle to track task progress. When search attempts fail to yield useful evidence, current single- and multi-agent systems can become trapped in repetitive loops, wasting search budgets and ultimately compromising the quality and completeness of the final output. We introduce SearchOS, a system-level multi-agent framework that turns fragile, implicit search progress into explicit, persistent, and shared state. First, we formulate open-domain information seeking as relational schema completion with grounded citations, where agents discover entities, populate attributes across linked tables, and anchor each value to source evidence. Then we design Search-Oriented Context Management (SOCM), which externalizes the evolving state into Frontier Task, an Evidence Graph, a Coverage Map, and Failure Memory. Built on SOCM, SearchOS applies a pipeline-parallel scheduling mechanism that overlaps the execution of sub-agents and continuously refills freed slots with tasks targeting unresolved coverage gaps to improve utilization and throughput. To schedule and control the execution of search agents, SearchOS introduces a Search Tool Middleware Harness that intercepts model and tool interactions to record grounded evidence and react to stalls or budget exhaustion, and provides a reusable hierarchical skill system comprising strategy and access skills to augment the agents' search process and avoid repeating failed search patterns across runs. On WideSearch and GISA, SearchOS leads all metrics among the evaluated single- and multi-agent baselines, paving the way toward robust information-seeking collaboration.

Retrieval systems are trained and evaluated on a static idea of usefulness: hand a document and a question to a reader model, see whether the answer improves, and score the document accordingly. The idea holds up when a document is read on its own. It breaks when a language model works as a search agent, issuing several queries and reasoning across turns, because a document can matter for what it lets the agent do next rather than for what it says about the current question. We measure that gap rather than argue it. Using a ReAct style agent over HotpotQA, we replay 1000 development questions and, for every document the agent read, delete it and re-run the rest of the trajectory from that point. Comparing the original run against its counterfactual gives a Counterfactual Trajectory Utility (CTU) score from three deltas: final answer quality, next query retrieval quality, and turn count. Crossing CTU against Static RAG Utility (SRU) over 23,322 document observations, the two are close to statistically independent (Spearman rho = -0.026). Roughly a third of the documents the agent reads are causally load bearing while looking useless to a static reader; we call these bridge documents. The pattern survives when the reader based axis is swapped for a BM25 and cross encoder proxy, giving a bridge cell of 27.2% on an evenly spread axis. A second experiment pins down the mechanism. Using the Observable Entity Relevance (OER) measure from prior work, entities that discriminate relevant from non-relevant candidates appear in the agent's next query 4.02 times more often than entities found only in non-relevant documents (6.1% vs 1.5%, n = 227,139). A bridge document earns its keep by handing the agent a discriminative entity that redirects the search. Static relevance and causal usefulness are different quantities in agentic retrieval, and optimizing the first does not deliver the second.

Evidence synthesis is crucial for turning primary research into reliable knowledge for science, medicine, education, and policy. Yet, quantitative evidence synthesis remains largely manual and difficult to scale. Here, we introduce AutoSynthesis, an end-to-end multi-agent system for automated meta-analysis. Given a research question in natural language, AutoSynthesis formulates a search strategy, retrieves scientific literature, screens candidate studies, assesses full-text eligibility, extracts quantitative statistics, computes standardized effect sizes, and finally performs random-effects meta-analysis. AutoSynthesis further supports heterogeneity analysis to examine how effect sizes vary across moderators, as well as risk-of-bias assessment. As output, AutoSynthesis produces a transparent report aligned with PRISMA guidelines. In our application, AutoSynthesis screened over 28 studies and extracted more than 20 quantitative claims. The pooled effect estimates produced by AutoSynthesis are similar to Hedges' $g$ of expert-conducted meta-analyses, indicating close agreement with manual evidence synthesis. Together, these results show that AutoSynthesis can make quantitative evidence synthesis more scalable, thereby supporting evidence-based decision-making across disciplines.

Healthcare multimodal AI must combine visual and textual evidence while remaining reliable and interpretable. Using MediaEval Medico 2025 as a retrospective GI endoscopy case study, we analyze design choices across nine documented systems for question answering and explanation quality. Parameter-efficient adaptation of pretrained backbones provides strong challenge performance, but answer-level gains do not consistently translate into faithful and complete clinical reasoning. Methods enforcing structured reasoning and explicit grounding show more reliable behavior across heterogeneous question types, although the evidence is correlational rather than ablation-based. These results motivate evaluation beyond lexical overlap, standardized evidence-linked explanations, leakage-aware data governance, and lightweight robustness and calibration checks. The findings support trustworthy multimodal healthcare AI based on data fusion, explainability, and resilient evaluation.

We develop data-driven algorithms for maintaining $N$ independent identical machines under a \textit{block replacement policy}, in which each machine is replaced upon failure and all machines are jointly replaced at regular intervals of length $k$. The goal is to learn the cost-minimizing interval $k^*$ from operational data when the lifetime distribution is unknown. At each decision epoch, the operator selects $k \in \{1, 2, \ldots, K\}$, observes the resulting failure history (a mixture of complete and right-censored lifetimes) and incurs a per-unit-time cost governed by the renewal function. We formulate this as a stochastic multi-armed bandit and propose Hoeffding- and Bernstein-based lower-confidence-bound algorithms achieving $O(K \log T)$ regret, matching the Lai--Robbins lower bound. Exploiting a nested observation property unique to block replacement, correlated variants attain $O((K-k^*)\log T)$ regret and require only $O(1)$ direct pulls of suboptimal arms $k < k^*$. A complementary Kaplan--Meier renewal algorithm estimates the lifetime distribution nonparametrically from censored data, achieving almost-sure policy consistency and empirically near-zero incremental regret at long horizons. We additionally analyze two average-cost MDPs: a time-elapsed formulation establishing that block replacement is optimal within its policy class for any lifetime distribution, and an age-vector formulation proving a monotone threshold structure under increasing failure rate distributions and providing a gold-standard cost benchmark. Numerical experiments confirm the theoretical ordering and reveal structural cost gaps between optimal block and age-dependent replacement.

Large language models (LLMs) increasingly serve as high-level planners for embodied agents, where linguistically benign instructions can become unsafe once grounded in the physical world. We study whether this physically grounded danger is the same safety problem as ordinary text-level content danger. Through hidden-state direction analysis and random-split null tests, we show that content danger (CD) and physical danger (PD) form separable signals in LLM representations across Qwen2.5-3B/7B/14B/32B, Phi-3.5 and SmolLM2. Building on the CD/PD separability, we propose PRISM, a single-layer L2-regularized logistic probe over full hidden states. PRISM achieves 86.2--87.7\% accuracy on SafeAgentBench with 11.7--13.7\% FPR, while same-scale LLM judges over-block safe tasks at 24.7--39.0\% FPR. We further introduce PhysicalSafetyBench-1K (PSB-1K), a contrastive benchmark of 1{,}000 physical-risk pairs without direct harm keywords, to test whether methods detect physically grounded danger rather than explicit unsafe wording. On PSB-1K, PRISM reaches 99.6\% accuracy and 0.7\% FPR, whereas a Qwen2.5-3B judge rejects 67.8\% of safe tasks. PRISM also replicates on SafeText and EARBench, supporting hidden-state probing as a representation-level method for physical safety beyond text moderation.

Multimodal large language models (MLLMs) often introduce errors when generating image captions, resulting in misaligned image-text pairs. Our work focuses on a class of captioning errors that we refer to as systematic misalignments, where a recurring error in MLLM-generated captions is closely associated with the presence of a specific visual feature in the paired image. Given a vision-language dataset with MLLM-generated captions, our aim in this work is to detect such errors, a task we refer to as systematic misalignment detection. As our first key contribution, we present Symbal, which utilizes a structured, dual-stage setup with off-the-shelf foundation models to identify systematic misalignments and summarize results in natural language. As our second key contribution, we introduce SymbalBench, a benchmark designed to evaluate automated methods on our proposed task. SymbalBench consists of 1.7 million image-text pairs from two domains (natural and medical images), organized into 420 vision-language datasets with annotated systematic misalignments. Symbal exhibits strong performance on this benchmark, correctly identifying systematic misalignments in 63.8% of datasets, a nearly 4x improvement over the closest baseline. We supplement our evaluations on SymbalBench with real-world evaluations, showing that (1) Symbal can accurately surface systematic misalignments in captions generated by four MLLMs and (2) Symbal is a powerful tool for auditing off-the-shelf image-caption datasets. Ultimately, our novel task, method, and benchmark can aid users with auditing MLLM-generated captions and identifying critical errors, without requiring access to the underlying MLLM. Code is available at https://github.com/Stanford-AIMI/Symbal.

World-action models (WAMs) are emerging as a promising foundation for embodied control: rather than predicting actions alone, they learn representations that couple action generation with future world prediction. This coupling is often viewed as a source of robustness, interpretability, and safety, as a robot's action can in principle be checked against its imagined future. In this paper, we show that this assumption is fragile. We introduce BadWAM, a unified framework for modeling and evaluating World-Action Drift Attacks: a new class of WAM-specific adversarial attacks that use small visual perturbations to break the alignment between what a WAM imagines and what it executes. BadWAM characterizes this attack surface along two natural criteria: attack strength and stealthiness. When the adversary prioritizes disruption, BadWAM instantiates an action-only adversarial attack, which directly drives the model toward task-failing actions. When the adversary additionally prioritizes stealth, BadWAM instantiates an imagination-preserving adversarial attack, which seeks to induce harmful action shifts while keeping the model's predicted future close to its clean imagination. Together, these two attacks capture a spectrum of WAM-specific failures: from overt action hijacking to stealthier cases where the model appears to imagine a plausible future but executes a desynchronized action. We evaluate BadWAM across different variants of WAMs. Results show that our attacks substantially reduce task success rates under closed-loop execution. For example, our action-only attack reduces the model performance from 96.5% to 43.1% success. The results of our imagination-preserving attack further exposes a WAM-specific vulnerability: moderate future-preserving regularization can maintain strong attack performance while reducing future imagination drift.

Real repository issues routinely include visual evidence such as screenshots, error dialogs, rendered UI states, and logs, yet repository-level issue localization is evaluated mostly as a text-only task. Existing multimodal SE benchmarks evaluate end-to-end repair, entangling localization with patch synthesis and obscuring whether visual input helped, hurt, or was ignored. We introduce \textbf{MM-IssueLoc}, a controlled benchmark and evaluation protocol for repository-level localization with visual evidence. MM-IssueLoc contains 652 issue-PR instances across 23 languages, with annotations for 7 image categories and 4 relevance levels. It provides file-level and function-level gold labels, paired text-only and with-image evaluation, and VCE-based diagnostics that convert images into structured textual evidence. We evaluate LLM-based and retrieval-based systems, including MM-IssueLoc-VL-Emb as a controlled multimodal retriever. Results show that existing systems remain far from reliable multimodal repository localization: the strongest agent reaches 38.96 file Acc@5 and 22.45 function Acc@10, while the strongest retriever reaches 33.86 function Acc@10. Cross-benchmark comparisons show that high localization scores on text-dominant SWE benchmarks do not transfer cleanly to multimodal issue localization. MM-IssueLoc turns visual evidence into an explicit evaluation variable, enabling future work to test whether systems improve by using visual evidence for localization, rather than by relying on text-only cues or downstream patch-generation effects.

Annotation quality is a major bottleneck in building reliable and explainable artificial intelligence (XAI) systems for mental health research. In depression-related datasets, labels are often assigned without structured evidence, symptom-level justification, or traceable alignment with the criteria of the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR), limiting both transparency and downstream model interpretability. We propose a self-evolving, expert-in-the-loop annotation framework for Major Depressive Disorder (MDD) that combines large language model (LLM)-assisted labeling with expert verification. The framework is intended to support the construction of explainable, DSM-5-TR-aligned datasets rather than to perform clinical diagnosis. It operates in three stages: candidate evidence selection from textual records, criterion-level DSM-5-TR analysis, and case-level synthesis that produces label-level diagnostic and severity annotations. A dual-memory architecture, composed of Example Memory and Reflection Memory, is designed to internalize expert feedback and iteratively improve future annotations without retraining. We describe this mechanism and leave its evaluation across multiple feedback cycles to future work. In addition to final labels, the framework exports clinical evidence, reasoning traces, and edit histories, enabling comprehensive auditability. In a pilot study using expert-reviewed samples, the proposed approach improves annotation consistency and explainability while reducing manual revision effort.

Reinforcement learning has proven effective for improving reasoning in large language models, but extending it to Masked Diffusion Language Models (MDLMs) remains challenging due to the intractability of the log-likelihood estimation. Existing approaches approximate this log-likelihood by modeling only the token predictions, ignoring the order in which positions are unmasked during generation. We observe that MDLM generation involves two decisions at each step: what tokens to place at each masked position and which positions to remask. We formalize this as a two-stage action MDP, showing that the policy gradient naturally decomposes into a token term and a masking term. Combining optimization of both terms leads to state-of-the-art outcomes on mathematical reasoning and coding benchmarks, with scores of 87.1% on GSM8K and 53.4% on MBPP.

Graphical user interface (GUI) automation remains challenging in real-world environments, where dynamic layouts, unexpected dialogs, and evolving interface states can cause autonomous agents to drift from user intent. Recent vision-based multimodal agents improve flexibility by operating directly over screenshots and natural language instructions, but planning and adaptation often remain internal, limiting users' ability to inspect, supervise, or correct system behavior. We present Plover, a plan-centric vision-based GUI automation system that externalizes task plans and replanning as persistent, inspectable, and revisable artifacts. Through a planner--executor architecture, Plover supports explicit supervision of evolving execution, localized correction through editable plans, natural-language guidance, and screenshot-grounded interventions, while preserving prior progress during repair. A formative study with six participants informed the interaction design. We then evaluate Plover through benchmark failure-case repair and scenario-based workflow analyses. Our results show that many autonomous GUI-agent failures are structurally repairable when plans remain visible and interventions are localized, and that explicit replanning helps make GUI automation more transparent, controllable, and adaptable.

AI benchmarks increasingly leverage item-level statistical models, particularly item response theory (IRT), to estimate model capabilities, rank systems, select informative examples, and diagnose benchmark quality. However, AI benchmark data often departs from the data regime of human testing, for which standard IRT estimation tools were originally developed: benchmarks typically involve fewer evaluated models, far more items, and capability distributions that may be skewed, clustered, or multimodal. We examine how these regime mismatches challenge the reliability of IRT modeling for AI evaluation. Using item parameters and capability distributions derived from six widely used LLM benchmarks, we simulate response matrices under three common IRT models and compare four estimation tools used in recent benchmark studies: marginal maximum likelihood, Markov chain Monte Carlo, variational inference, and a neural pseudo-Siamese estimator. Across 18,000 simulation conditions, we systematically evaluate computational feasibility, scalability, and the reliability of IRT inferences about model rankings, predicted performance, and item characteristics. Results show that classical estimators can become infeasible in large benchmark settings, whereas scalable estimators can produce unreliable item-level and ranking inferences with small or nonnormally distributed model sets. This study identifies when latent trait models reliably support or risk distorting AI benchmarking claims, and what sample sizes and diagnostics are needed for trustworthy use.

Automated fulfillment warehouses must continuously assign and execute pickup-and-delivery work while avoiding congestion. In many-to-many Multi-Agent Pickup and Delivery (MAPD), a request specifies a stock-keeping unit rather than fixed endpoints, requiring the controller to select an agent, source, and destination before path planning. Existing graph-guidance methods primarily influence routing after goals are fixed, leaving endpoint instantiation uninformed by recent traffic. We introduce Stigmergic Graph Memory (SGM), a bounded, decaying memory layer that records recent execution signals on warehouse nodes and directed edges to rank feasible endpoints and route preferences without altering collision constraints or planner validity. Across paired request streams on five layouts, three load levels, and 25 seeds per condition, SGM outperforms two reconstructed many-to-many allocation baselines in all 15 map-load conditions, with paired throughput gains of 20.5-36.7%. These results show that recent execution memory can improve warehouse throughput by shaping which feasible goals enter the planner, not only how agents travel to already fixed goals.

Ordinary differential equations (ODEs) are widely used to model dynamical systems in physics, biology, neuroscience, and physiology, but in many applications some equations of the dynamics are unknown and only a subset of the state variables are measured. We propose a hybrid neural--physics framework in which the known components of the ODE are kept explicit and the missing components are represented by a neural network. The proposed method consists of two stages where we alternate between state and parameter estimation and iterate until a predetermined criterion is met. Specifically, in the first step, we treat the model parameters as being known and we infer the latent states from the available measurements using a Rauch--Tung--Striebel (RTS) smoother. In the second stage, we treat the smoothed trajectories as being known and use them to estimate the neural networks' parameters through backpropagation. We evaluate the method on benchmark systems spanning linear, nonlinear, and stiff dynamics under partial state observation. Across these settings, the proposed method learns missing ODE components from incomplete measurements while exploiting and retaining interpretable mechanistic structure and improving latent-state reconstruction and long-horizon prediction.

Multimodal large language models (MLLMs) are increasingly used to interpret visualizations, yet current evaluations remain largely chart-centric and provide limited evidence of understanding of scientific visualization (SciVis). We benchmark six MLLMs on the scientific visualization literacy assessment test, a standardized SciVis literacy assessment comprising 49 items based on 18 scientific visualizations and illustrations, spanning 8 techniques and 11 task types. We evaluate three closed-source and three open-source models under a closed-world protocol and compare their performance using data from 485 human participants. Results show that current MLLMs do not exhibit uniform SciVis literacy. Gemini is the strongest model overall, exceeding the human mean across the evaluated subsets, whereas the open-source models remain below the human baseline. Performance is highly uneven across techniques and tasks: models perform best on scientific illustration, search, and spatial understanding, but struggle on texture-based and integration-based visualizations and on quantitative estimation. Error analysis reveals recurring failures in fine-grained quantitative estimation, flow-direction interpretation, and grounded encoding interpretation. These findings position SciVis literacy as a necessary benchmark dimension for evaluating multimodal AI systems. Our code and model outputs are publicly available at https://github.com/patdmp/mllm-scivis-lit-benchmark.

Most medical AI benchmarks measure whether a model knows the correct answer. MedFailBench asks a different question: which safety boundary failed? We present a clinician-built synthetic benchmark and failure atlas that labels medical AI errors by severity (1--5) and safety gate type (missed urgent escalation, unsafe remote dosing, unsafe discharge reassurance, evidence fabrication, unsafe protocol execution, source support gap). The current public release (v0.2.1) contains 44 clinician-reviewed synthetic cases with severity annotations, a live HuggingFace leaderboard preview, a safety gate taxonomy, a clinical severity rubric, and an automated pipeline for archiving model-response screening runs. No patient data, clinical validation claims, or model rankings are included. MedFailBench is released under Apache-2.0 and CC-BY-4.0 and carries the Zenodo DOI 10.5281/zenodo.21205535.

Humanoid control requires natural whole-body coordination, precise real-time responses to control signals, and robust generalization across diverse environmental contexts, making it a cornerstone for generalist embodied agents. Behavior Foundation Models (BFMs) have recently emerged as a promising solution to address these challenges by leveraging large-scale behavioral data to achieve superior expressiveness, versatility and generalization. However, despite growing interest in scaling BFMs to further improve their capabilities, it remains unclear how key factors, including the learning paradigm, behavioral data and model architecture should be coordinated to enable effective scaling. In this work, we revisit the scaling recipe for BFMs and demonstrate that substantial performance gains can be achieved through the coordination of three core components: 1) the learning paradigm of motion tracking that reformulates diverse humanoid control problems as the reproduction of integrated whole-body behaviors in the global frame; 2) the strategic synergy between on-policy rollout quantity and reference motion diversity; and 3) the expressive and scalable model architecture termed Humanoid Transformer that facilitates the natural emergence of structured behavioral representations. Through extensive experiments in both simulation and real-world deployment, we demonstrate that our approach yields significant improvements in control fidelity and task generalization, reducing Mean Per-Keypoint Position Error (MPKPE) on the test set by over 10% in local mode and 82% in global mode compared with existing humanoid controllers. These results establish BFM as a principled and effective foundation for scalable and general-purpose humanoid control.

On-policy distillation is an alternative post-training method in reinforcement learning that alleviates the constraints imposed by reward models by providing token-level supervision from a teacher model. Although on-policy distillation has been studied and applied across various settings, its fundamental design remains underexplored. In this paper, we introduce a new distillation reward, termed the delta signal, instead of directly imitating the teacher's output distribution. The delta signal is defined as the difference between the teacher model and its base model prior to instruction tuning for reasoning capability. It therefore captures the changes induced by reasoning tuning and provides a more direct signal for transferring reasoning capabilities. Using extensive empirical evidence, we show that the delta signal substantially improves on-policy distillation and refer to the new distillation method as On-Policy Delta Distillation (OPD$^2$). Experiments across mathematics, science, and code-reasoning benchmarks demonstrate that OPD$^2$ consistently outperforms conventional on-policy distillation, enabling reasoning LLMs to achieve strong performance with only a short post-training period. Code will be available at https://github.com/naver-ai/opd2

World models are usually evaluated as components of model-based reinforcement learning (MBRL) systems, while the world models themselves are rarely studied in isolation. We examine five representative visual world-model agents in Atari Pong: DreamerV3, DIAMOND, TWISTER, Simulus, and STORM. After reproducing their training pipelines and matching the reported agent performance, we freeze the learned world models and evaluate them with a closed-loop rollout diagnostic: a policy trained separately from the corresponding MBRL agent interacts with each frozen model, and the generated video trajectories are inspected for visual and dynamical errors. Across all five models, the rollouts contain clear failures, including ball disappearance, incorrect ball motion, and invalid ball-paddle interactions. Beyond visual trajectories, we further evaluate them with pixel-space zero-shot MBRL, where a new policy is trained entirely inside a frozen world model and then evaluated in the real environment. Across all five models, the resulting policies substantially underperform those produced by the corresponding original MBRL training pipelines. The gap is particularly large for DreamerV3, whose mean return drops from -5.5 to -20.9, near the minimum Pong return of -21. We hypothesize that insufficient modeling of task-critical concepts, such as the ball in Pong, may contribute to these failures. We therefore propose Concept-Guided Spatial Regularization (CGSReg), an auxiliary pixel reconstruction loss applied to segmented concept regions. Experiments show that CGSReg improves both closed-loop rollouts and pixel-space zero-shot MBRL in DreamerV3, DIAMOND, and TWISTER. Its effects vary across the remaining models and evaluation metrics, indicating that CGSReg alone does not address all world-model bottlenecks.

Recent FPGAs have improved deep learning (DL) inference efficiency through dedicated tensor blocks and in-BRAM computation. ReRAM-based analog in-memory computing (IMC) pushes efficiency further, offering an order-of-magnitude improvement in compute density and energy efficiency over conventional digital logic by performing vector-matrix multiplication (VMM) directly within the ReRAM crossbar; prior work has integrated such IMC blocks into FPGAs for DL inference. However, conventional IMC designs support only static-weight VMM, leaving nonlinear operations and dynamic matrix-matrix multiplication (DIMM) to the FPGA fabric. As a result, the benefits of IMC are largely confined to static-weight models, whereas Transformer-based models, which rely on frequent nonlinear and DIMM operations, gain only limited improvement. Moreover, the ADCs within each IMC block consume more than 70% of its area and power, further limiting system efficiency and scalability. To address these limitations, we propose a novel FPGA architecture that integrates an ADC-free IMC block, replacing the conventional ADC with analog content-addressable memories (ACAMs) that natively perform nonlinear operations inside the block. To fully exploit this block, we conduct an FPGA-aware design-space exploration that determines optimal crossbar dimensions while balancing FPGA area, flexibility, and DL performance, and we develop an efficient mapping that leverages ACAMs to carry out DIMM operations, extending the applicability of IMC to attention computation. On CNN and Transformer-based benchmarks, the proposed architecture achieves up to 40x and 1.9x higher energy efficiency and 4.1x and 2.5x higher area efficiency, respectively. Overall, it significantly improves FPGA DL inference efficiency and sustains robust gains on Transformer-based workloads across long input sequences, advancing domain-specialized FPGA design.

Parameter-efficient fine-tuning reduces model and optimizer memory, but dense attention still makes long training sequences expensive. We combine Hierarchical Global Attention (HGA) with segment-wise backpropagation and tiered KV storage. Only the active segment remains differentiable in VRAM; older KV is detached into RAM or NVMe, and HGA loads a bounded set of exact historical tokens for each query block. On Qwen3-8B with 4-bit QLoRA and PG19, dense training on a 16 GB Quadro RTX 5000 fits 2,048 tokens but fails at 4,096, whereas HGA reaches 16,384 tokens with 15.28 GB peak VRAM. Under evaluation the same adapter runs through 131,072 tokens on this card; VRAM is not constant but grows gently with the resident chunk summaries, so RAM and NVMe capacity set the practical limit beyond these lengths. At the shared 2K training length, HGA-trained and dense-trained adapters obtain 2.7405 and 2.7383 nat under the same dense-attention readout, while the stock model obtains 2.9541. At this boundary HGA training is already marginally faster (217.75 vs. 207.02 tokens/s), and the HGA-to-dense throughput ratio improves from 1K to 2K; because HGA keeps the attended historical set per token approximately constant while dense work per token grows, we expect this lead to widen as context grows. Dense attention is used for the main quality and retrieval comparisons so that they measure the learned weights and remain compatible with standard generation frameworks. HGA can also be used for retrieval and generation; an optimized production-grade serving implementation is under development.

The formation of political coalitions is a complex negotiation driven by both concrete policy objectives and deep-seated ideological convictions. While Large Language Models (LLMs) open new avenues for computational political science, the neutrality and helpfulness biases instilled by Reinforcement Learning from Human Feedback (RLHF) prevent them from sustaining steadfast partisan behaviour. We present a multi-agent framework that reconciles factual grounding with ideological alignment by combining Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Retrieval-Augmented Generation (RAG): DPO instils aggressive party-specific personas, while a per-party RAG pipeline keeps each agent bounded to its official manifesto. We operationalize the framework on the 2019 Flemish election, deploying the partisan agents in a hub-and-spoke negotiation arbitrated by a formateur. To make the emergent negotiation interpretable, we introduce a Multi-Layered Information Lineage Topology (MILT) that traces every clause in the final agreement back to its manifesto origin and classifies it into five provenance states, a Coalition Influence Score (CIS) that aggregates these traceable contributions to identify which party shaped the agreement, and a real-world grounding pass that benchmarks each simulated provision against the historically adopted coalition agreement. Across three independent simulations the framework yields a stable winner and ranking (N-VA ahead of CD\&V and Open Vld), and manifesto-anchored lineage reliably predicts real-world materialization whereas hallucinated content does not. The result is a transparent, scalable testbed for the ex-ante exploration of party compatibility and formateur-mediated compromise.

Multimodal models such as CLIP learn a shared embedding space for cross-modal retrieval, but continual adaptation to sequentially arriving data can disrupt the cross-modal alignment acquired from earlier phases. Conventional continual-learning methods return a single checkpoint, which commits every retrieval direction to the same stability-plasticity trade-off. We propose AlphaWiSE, a post-hoc weight-space interpolation method that composes two frozen source checkpoints. For each aligned parameter tensor identified by its checkpoint key, AlphaWiSE fits one scalar interpolation coefficient shared by all tensor entries. The coefficients are fitted on a smaller exemplar memory and used to materialize one interpolated checkpoint. The deployed model has the same architecture and parameter count as either source checkpoint, which does not require additional inference time. Extensive experiments on audio-image-text retrieval show consistent improvements over strong continual-learning baselines across multiple retrieval directions and evaluation metrics.

Rubrics provide structured, fine-grained signals for training and evaluating large language models (LLMs). Yet reliable query-specific rubrics are difficult to construct. Existing approaches often derive supervision from human-written rubrics, preference data, or sampled responses. Direct query-to-rubric generation avoids these resources, but provides no explicit check that a plausible rubric is useful. Such a rubric may fail to distinguish answer quality, reward an optional style, or penalize a valid alternative strategy. We introduce Rubrics on Trial, a query-only framework that evolves a rubric set from an empty set without external annotations or model training. It derives supervision solely from synthetic rubric-conditioned response pairs and validates each proposed rubric before adding it, screening out non-discriminative, over-specific, and style-only candidate rubrics. Experiments across five preference benchmark suites demonstrate the effectiveness of Rubrics on Trial, which achieves the best average accuracy and leads on six of seven evaluation sets.

Understanding the brain increasingly depends on integrating evidence across scales, modalities, and disciplines. Addressing a single research question therefore requires a coordinated sequence of operations, from surveying prior work to executing analyses and interpreting results in light of domain knowledge. AI agents promise to accelerate this process, but current agents lack domain expertise in brain science, may fabricate claims, drift during multi-step reasoning, and offer few defined points for expert intervention. These failures are especially costly in brain science, where conclusions feed into downstream scientific claims and depend on laboratory-specific expertise and careful human judgment. We present \textbf{BrainPilot} a \textbf{fully open-source} multi-agent system that accelerates brain science research with traceable logs and agent-verified results. A principal investigator (PI) agent coordinates specialist agents grounded in curated domain knowledge: a unified brain science knowledge base containing 7{,}233 indexed items and a skill library of 72 reusable methodology units across seven research domains. Every major step is recorded in the Graph of Trace, an auditable record that links subgoals, tool use, evidence, and claims and allows researchers to follow and inspect the workflow. An Auditor agent further integrates fabrication checking into the workflow. For evaluation, we run three brain science tasks from Agents' Last Exam, introduce our own benchmark, \textbf{BrainPilotBench-v0}, and present additional end-to-end case studies. Across these evaluations, BrainPilot with an open-source backbone model attains performance comparable to state-of-the-art agent framework with less costs.

Many engineering problems involve phenomena whose governing equations are poorly characterized or only partially known. Surrogate modeling techniques such as neural networks can capture the behavior of these systems, but they typically demand large training datasets that are difficult to obtain in engineering contexts and yield models with limited physical interpretability. The Sparse Identification of Nonlinear Dynamics (SINDy) method addresses both limitations by performing sparse regression over libraries of candidate nonlinear terms, recovering interpretable governing equations from comparatively small datasets. Although SINDy has been demonstrated extensively on canonical benchmark systems, its application to practical engineering problems is less widely documented. This tutorial introduces the SINDy method and progressively builds toward its main extensions, from noise-robust weak-form and ensembling-based variants to constrained and parametrizable formulations. The paper and the accompanying tutorial (available at https://github.com/paullililili/SINDy4Engineers) is organized in three parts: the first introduces the standard SINDy algorithm and progressively extends it, inviting readers without prior knowledge to follow each step and adapt the methods to their own problems; the remaining two parts present detailed case studies on (1) the system identification of an unmanned aerial vehicle and (2) a chaotic thermosyphon heat exchanger. Through these examples, we aim to demonstrate that SINDy is simple to implement yet flexible enough to serve as a valuable identification tool for advanced engineering applications.

This article presents a novel estimator for performing off-policy evaluation using only offline data for contextual bandits. The proposed estimator, Kernel-WIS is demonstrated to be asymptotically consistent and to empirically outperform strong baselines (including vanilla weighted importance sampling), particularly under complex conditions including behaviour policy miss-specification. The benefit of Kernel-WIS is derived from combining the bounded property of vanilla weighted importance sampling with the linearity of vanilla importance sampling.

Predictive world models enable robots to plan by imagining the outcomes of their actions, but their value for control hinges on generating many rollouts quickly. This creates a bottleneck for diffusion-based world models: multistep sampling makes each rollout expensive, limiting large-scale action search at inference time. We introduce DriftWorld, an action-conditioned world model based on drifting generative models. Rather than denoising iteratively at inference, DriftWorld learns an action-conditioned drift during training, allowing it to generate future frames from the current observation and a candidate action sequence in a single forward pass at 30+ fps, which is 17x faster on average than diffusion based baselines. We evaluate DriftWorld on standard vision-based robotic manipulation benchmarks, including Bridge-V2, RT-1, Language Table, Push-T, and Robomimic. By producing rollouts that are both accurate and fast, DriftWorld achieves state-of-the-art decision-making performance with far less inference time than diffusion-based world model baselines. Beyond online control, DriftWorld can also serve as an offline simulator for ranking real-world robot policies, with rollout-based scores correlating with ground truth at up to 0.99. These results show that drifting models are a strong fit for robot world modeling, where fast, high-quality imagination directly supports planning and policy evaluation.

The Internet taught us that the value of a network depends on \emph{how} its nodes connect: broadcast stars scale as $V\!\propto\!N$ (Sarnoff), fully-connected meshes as $N^2$ (Metcalfe), and group-forming networks as $2^{N}$ (Reed). We ask the analogous question for networks of AI agents. We model the net value of connection as a function of coordination-group size, derive from it the properties an optimal collaboration protocol must have, and introduce ANet Patu-1 -- a self-organizing consensus protocol in which the network continuously re-forms its own coalitions, adaptively riding the upper envelope of all three regimes at $O(1)$ parallel consensus rounds. To measure value without opinion-grading, we score an emergent protocol by formally specifying it and deriving its complexity, the way distributed algorithms are analyzed. Two results follow. (i)~Emergence -- a crowd of the \emph{cheapest} model, when heterogeneous, starts weak but its collective value compounds with $N$ and \emph{overtakes} a crowd of a far \emph{stronger} model that is homogeneous: a crossover that marks a scaling law for collaboration rather than for scale. (ii)~Reflexivity -- a heterogeneous network, given only its own problem and no design hints, converges on ANet Patu-1 itself, reconstructing the high-dimensional law that governs its own connective value.

Lattice quantum chromodynamics (LQCD) provides a first-principles framework for computing hadronic observables, but its practical use remains limited by the substantial expertise required to turn research motivation into reliable computing workflows. Here we present \textsc{LQCDMaster}, a tool-augmented, skill-guided and domain-specialized scientific computing agent that converts natural-language LQCD research tasks into executable PyQUDA computing workflows, including measurement scripts, job-submission artifacts, execution logs and numerical outputs. The system combines agentic planning, expert-annotated LQCD skills and a deterministic Wick-contraction tool to constrain the algebraically fragile components of code generation. We evaluate \textsc{LQCDMaster} on a benchmark at the forefront of scientific research, comprising 70 LQCD computing tasks, with observables covering local and nonlocal two-point functions, Wilson loops, meson and baryon three-point functions. The generated workflows exactly reproduce expert-written implementations in 63 of 70 tasks at machine precision, with three additional discrepancies attributable to convention mismatches. Across representative observables, the agent reduces implementation time from hours to minutes while preserving end-to-end numerical validation. Further, we present a typical case of \textsc{LQCDMaster}-driven exploration: a lattice computation of light-cone distribution amplitudes with diagonal Wilson-line, a quantity accessible with standard methods but never before computed, and computation of the spectrum of proton, deuteron, triton, hyperon, hyperdeuteron and hypertriton. This work pioneers the paradigm of agentic scientific computing by automating the end-to-end scientific computing workflows in lattice QCD research, lowering its barrier and facilitating the exploration and verification of non-standard scientific ideas.

Large language models are increasingly evolving from text generators into general agents capable of understanding user requests, invoking external tools, and completing complex tasks through interaction. However, existing agent benchmarks often focus on limited scenarios, tool ecosystems, or interaction formats, making it difficult to systematically characterize model capabilities across heterogeneous application settings. We introduce OmniaBench, a benchmark for evaluating general agents across diverse scenarios with explicit state spaces. We derive application-oriented scenario knowledge from app stores, product documents, industry resources, Web retrieval, and human refinement, forming a hierarchical taxonomy that spans ToC, ToB and ToE with 90 level-1 and 354 level-2 domains. Based on this taxonomy, we construct executable environments and synthesize single-turn and multi-turn tasks through four complementary routes: DAG, DAG-S, Solver, and Program. OmniaBench further introduces a ten-dimensional capability taxonomy and eight compositional atomic difficulty factors to support fine-grained evaluation and analysis. The resulting dataset contains 1,431 tasks, together with a challenging subset of 644 tasks designed to reduce evaluation cost and mitigate potential contamination of the full set after public release. The bench presents substantial challenges to current frontier models, with even Claude-Sonnet-5 and GPT-5.6-Sol achieving Overall Pass@1 scores of only 58.54 and 57.14, respectively. Further analyses reveal clear differences across domains and capabilities, as well as persistent limitations in planning, constraint maintenance, and adaptive correction. OmniaBench provides a broad and diagnostic benchmark for characterizing the capability boundaries of general agents.

Channel foundation models (CFMs) are developing rapidly, with recent studies reporting benefits from pretraining across downstream wireless tasks. Yet CFMs are commonly evaluated in model-specific pipelines with different data, radio configurations, partitions, adaptation procedures, task definitions, and metrics. Reported comparisons therefore tend to show that pretraining improves over supervised training from scratch within one pipeline, but neither rank CFMs nor compare them fairly with task-specific models. We release CFM-Bench, a unified multi-domain, multi-task benchmark designed to address this gap. It curates six channel configurations spanning 3GPP statistical simulation, two independent ray-tracing pipelines, industrial and aerial measurements, and synchronized vehicular multimodal simulation. Official partitions isolate complete trajectories, measurement sessions, vehicle links, simulation realizations, or buffered spatial regions. CFM-Bench does not prescribe an external pretraining corpus or strategy; no benchmark split may be used for foundation-model pretraining, and the official training split is reserved exclusively for downstream fine-tuning. The benchmark additionally requires disclosure of all data used during model development and prohibits training-stage use of official test units. Six task groups are organized along three CFM application dimensions: physical-layer (PHY) channel intelligence, radio-access-network (RAN) decision intelligence, and integrated sensing and communication (ISAC). They cover CSI feedback, frequency and temporal channel extrapolation, propagation-state classification, current- and future-beam prediction, and single-frame and temporal localization. CFM-Bench provides a common substrate for comparing the transferability of channel representations across models, domains, and tasks.

Most existing approaches to AI-Generated Text Detection (AIGTD) treat documents as static objects and base their decisions on aggregate statistics or globally compressed embeddings. However, this perspective overlooks the inherently dynamic nature of autoregressive generation, where content evolves progressively through the latent space. In this paper, we reformulate AIGTD as the problem of distinguishing between latent generation trajectories. Instead of relying on static representations, we model how textual representations evolve across the sequence. To this end, we propose Geometric Trajectory and Contrastive Learning (GTCL), a framework that segments the document into ordered local units, encodes each unit in an embedding space, and constructs a structured and sequence-level representation. GTCL then applies contrastive learning to these trajectories to learn geometric regularities associated with the autoregressive generation. Evaluations performed on three different benchmarks and several approaches show that GTCL outperforms detection baselines consistently, which implies that explicitly modeling sequential dynamics provides robust discriminative signals across models and domains. These results suggest that modeling trajectory differences could improve detection and open up a dynamic direction that has been underexplored in previous AIGTD literature.

Text-to-image (T2I) models can synthesize realistic, prompt-aligned images, yet samples generated for the same prompt often cover only a small subset of visually distinct modes. This limits the diversity of images, and for person-centric prompts, can reflect or amplify demographic skew. We formalize this problem as coverage of a predefined set of semantically specified modes, which we call target-mode coverage. We then propose multi-axis max@K, a group-based reinforcement learning objective for improving such coverage in diffusion-based T2I models. Given a group of samples and one score per target category, multi-axis max@K first takes the maximum score across samples for each category and then sums these category-wise maxima. The resulting credit assignment gives a sample positive weight on a category only when it increases that category's group-wise maximum, allowing different samples to contribute to different categories. We first validate the credit-assignment mechanism on a synthetic mixture and on SD3.5-M using deterministic pixel-based color rewards. We then evaluate the same objective on perceived-appearance fairness. Across three automatic evaluators on held-out prompts, multi-axis max@K improves the Fairness Score by 0.23-0.36 relative to the base model, while maintaining image quality and text alignment.

A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at deployment. The gap is especially important for AI agents, whose observations, tool outputs, documents, and prior decisions accumulate over long trajectories. LongStraw is an architecture-aware execution stack for million-token RL post-training under a fixed GPU budget, instantiated with Group Relative Policy Optimization (GRPO). It evaluates the shared prompt without autograd, retains only model-specific state needed by later tokens, and replays short response branches one at a time, reducing the live training graph at the cost of additional replay time. We implement it for the hybrid recurrent and full-attention Qwen3.6-27B and the compressed-attention mixture-of-experts GLM-5.2. On eight H20 GPUs, LongStraw completes grouped Qwen scoring and response backward at 2.1M positions for groups of 2 and 8; increasing the group size adds only 0.21 GB of peak allocated memory, while a separate stress test reaches 4.46M positions. On 32 H20 GPUs, we validate the end-to-end LongStraw execution path for a 2.1M-token prompt across all 78 layers of GLM-5.2. These experiments establish execution capacity rather than complete training correctness because the captured prompt state is detached and some distributed forward and gradient composition paths remain incomplete.

Modern generative models are increasingly trained using model-generated signals, creating both opportunities for self-improvement and risks of collapse. We study optimal self-distillation (SD) for rectified flow (RF): given a suboptimal teacher velocity field, can a student trained on a mixture of true RF velocities and teacher velocities provably improve the teacher? For linear RF with ridge regularization on fixed interpolation pairs, we prove an exact affine path identity, derive the optimal mixing coefficient in closed form, and show strict improvement in integrated velocity risk whenever the teacher risk is nonstationary along the regularization path. The optimal coefficient obeys a sign rule: positive mixing corrects under-regularized teachers, while negative mixing corrects over-regularized teachers. We also give one-shot generalized cross-validation (GCV) and validation tuning procedure that avoids grid search over mixing weights and repeated refitting. Combining this theorem with RF Wasserstein convergence bounds, we show that optimal self-distillation improves the velocity estimation terms controlling continuous-time and finite-step generation error. Experiments with Gaussian models, Gaussian mixtures, and image data show that optimal self-distillation improves velocity risk, mode recovery, and finite-step generation relative to both the teacher and pure distillation.

World Action Models (WAMs) enable semantically- and physically-informed control but are brittle under distribution shift. In this work, we use mechanistic interpretability to study how robustness-relevant perturbations are represented in WAM activation space. Comparing activations across successful and unsuccessful rollouts, we find some WAM architectures exhibit low-dimensional linear separability for robustness-critical features, while others do not. This motivates the use of contrastive activation directions for training-free WAM steering. We also show that local linearity in WAM activation dynamics enables efficient feedback steering via model-based optimal control, yielding World-Action Linear Quadratic Regulator (WA-LQR), a minimally-invasive reduced-order LQR controller. Via mechanistic evaluations, we predict strong steerability in the Cosmos-Policy and DiT4DiT models but weak steerability in LingBot-VA, consistent with steering intervention results. On Cosmos-Policy and DiT4DiT, WA-LQR generalizes contrastive directions to new tasks and improves robustness to camera, gripper, and visual-noise perturbations over unsteered and prompt steering baselines.

Synthetic face datasets have become effective enough to train face recognition models with accuracy rivaling that of models trained on real photographs. This progress sidesteps the ethical and legal burdens of collecting real biometric data, yet evaluation has not kept pace. Even studies that train entirely on synthetic images still rely on real-face benchmarks to measure performance, leaving the privacy problem only half solved. We ask whether synthetic datasets can replace real benchmarks for face recognition evaluation. We test 12 synthetic datasets against 7 established real benchmarks using 24 pre-trained models that span both convolutional and transformer architectures. Our evaluation covers biometric verification metrics, similarity score distributions, cross-model ranking consistency, and the underlying distributional properties of each dataset. Benchmarking fidelity varies widely across the synthetic candidates, but the two strongest, MorphFace and Vec2Face, reproduce the relative behavior of real benchmarks and reach agreement levels that fall within the natural disagreement already observed among the real benchmarks themselves. These results establish that well-constructed synthetic datasets can support reliable comparative evaluation for face recognition, moving the field closer to a fully synthetic and privacy-preserving pipeline for both training and benchmarking.

Real-time video generation demands fast decoding as much as fast denoising, yet current latent video diffusion models rely on 3D convolutional decoders that are slow and memory-intensive at high resolutions or for long video. We introduce FlashDecoder, a fast, memory-efficient pure-Transformer video decoder that decodes latents to pixels frame by frame. At each step, the current frame attends only to a fixed-size window of past frames through a rolling KV cache. The fixed temporal window keeps decoding fast and memory bounded regardless of video length, enabling constant-latency streaming. Because frames are processed sequentially, temporal causality is enforced without explicit attention masks, enabling training at resolutions up to 1080p and matching the reconstruction quality of convolutional decoders. On the Wan2.1 and Wan2.2 latent spaces, FlashDecoder matches each convolutional decoder in reconstruction quality (e.g., 41.55dB vs. 41.49dB PSNR at 1080p) while decoding 3.6x-4.7x faster with up to 11x less memory on a single H100 GPU. With architecture-aware inference optimizations, the speedup widens to 12x.

Addressing a structural-engineering request requires more than a single answer; it requires a chain of interdependent artifacts: interpreted requirements, a computable model, validation records, solver outputs, code-check records, and a final report. Evaluations centered on question answering or script generation rarely verify this complete evidence chain and may therefore reward fluent outputs even when the underlying engineering workflow is incomplete, internally inconsistent, or non-executable. To address this limitation, we present StructureClaw, an artifact-centered workbench in which LLM agents operate through governed engineering skills, typed tools, shared artifact state, and local analysis backends. We also introduce StructureClaw-Bench, an executable benchmark of 150 controlled scenarios spanning standard workflow execution, interactive robustness, and multimodal structural-model reconstruction. A scenario succeeds only when all required artifact- and execution-level assertions pass in a single run. Across ten agent-model configurations, each evaluated on the same 50 standard cases, the average Success Rate rises from 56.8% with the generic-skill baseline to 88.6% with the full automatic workflow. The interactive and multimodal evaluations identify two prominent remaining challenges: safe handling of invalid numerical inputs and fixture-consistent reconstruction of structural models. These findings show that artifact-centered evaluation can expose workflow-level failures that are difficult to identify from final responses alone, providing a more rigorous basis for evaluating and improving structural-engineering agents. The code and benchmark are available at https://github.com/structureclaw/structureclaw.

Autonomous coding agents increasingly execute multi-step software work, but lifecycle states such as reviewed, tested, DONE, and ready-to-merge remain claims unless supported by current evidence. We present Proof-or-Stop Lifecycle Control, a method that permits lifecycle transitions only when fresh, tracked-source-state-bound, mechanically verifiable evidence satisfies the relevant gate. The method treats agent outputs as claims rather than lifecycle state, and uses proof operationally to mean gate-admissible evidence under a stated trust model, not semantic program correctness. We evaluate an open-source implementation through mechanism tests, a powered control-policy ablation, and operated self-application evidence. The unattended-loop engine passed 10 of 10 scenarios with zero false-DONE, and local-key receipt bundles rejected 18 tamper classes with zero false accepts. In a 9,240-cell ablation, the pre-registered A4 versus A2-prime comparison reduced visible-pass/hidden-fail amplification from 31 of 1,800 injected cells under a compute-budgeted naive loop to 2 of 1,800 under the gated loop, a 1.6 percentage-point improvement in not-amplified rate with a 95 percent confidence interval of [0.8, 2.5]. A near-compute A3 versus A4 comparison, 14 of 1,800 versus 2 of 1,800, indicates that the gain is associated with enforcing review as a lifecycle gate rather than merely adding a reviewer. The self-application corpus contains 565 stories and 1,007 review findings, with 94.8 percent resolved, plus a 68-row high/critical cross-vendor exhibit. These results support Proof-or-Stop as a model-agnostic, host-neutral control layer for deciding which autonomous-agent claims a lifecycle may act on. The evaluation is limited to one model family, 24 ablation tasks, and a self-hosted corpus.

Finetuning language models on small, curated datasets is standard practice for adapting them to specific policies or domains. We show that finetuning on narrow, factually-defensible, moderation-passing data can cause broad ideological shifts across unrelated domains, while preserving general capabilities. Training GPT-4.1 on right- or left-leaning economics Q&A yields matched ideological shifts on topics such as criminal justice, the environment, and cultural taste. The same effect appears with plausibly-deployed datasets such as workplace HR policy and practical finance queries, as well as on a science-pseudoscience axis where food-safety finetuning increases sycophantic agreement with users expressing false health beliefs. We call this phenomenon ideological generalisation and propose a methodology to measure two properties: breadth, how far the shift reaches across topics absent from training, and amplification, how much finetuning intensifies the shift relative to few-shot prompting on the same examples. We show that few-shot prompting indicates the direction of generalisation but finetuning pushes the model to further extremes, including to far out-of-distribution outputs such as endorsements of race-IQ connections and political violence. The effect replicates on Gemma-3, holds under judge-free evaluations and external benchmarks, survives mixing with generic data, and leaves GSM8K accuracy within $\pm 1$pp of the baseline.

Temporal Knowledge Graph (TKG) reasoning under the extrapolation setting focuses on forecasting future time-stamped events (facts) from historical data in a temporal knowledge graph. Existing approaches, reinforcement learning (RL)-based multi-hop reasoning methods are prominent for TKG reasoning because they produce human-interpretable predictions via explicit multi-hop path tracing. However, during RL training, rewards are typically sparse, and exploration is highly inefficient due to the vast, time-evolving action space. These issues hinder efficient training and often limit overall performance. To address these challenges, we propose RAPTOR (Reachability-Aware Pretraining for Efficient Target-Oriented Path Exploration), a self-supervised pretraining method that injects a reachability-aware inductive bias to the agent. By learning to estimate the reachability of candidate actions to the target entity, RAPTOR reduces exploration over unpromising paths and provides a strong initialization for downstream RL fine-tuning. Experimental results on the ICEWS14, ICEWS05-15, and ICEWS18 datasets demonstrate that RAPTOR pretraining markedly improves the training efficiency and consistently outperforms conventional baselines, establishing it as an effective approach for enhancing RL-based multi-hop reasoning methods for TKG reasoning.

With a large controlled vocabulary as the label set, the task of automated subject indexing in a library can be understood as a multi-label classification task. If the set of subject terms is large, the problem fits the Extreme Multi-Label Classification (XMLC) objective. In this study, we apply a selection of specialised supervised XMLC methods to the test case of subject indexing contemporary German scientific literature, collected at the German National Library (DNB). We contrast these results by including a classical lexical matching baseline and three of our own recently developed LLM-based methods into the benchmark. Algorithms are evaluated and compared in several metrics. This includes binary relevance comparisons with previously indexed material, as well as graded relevance ratings by professional subject librarians. A challenge for all methods is to reliably make suggestions from the long tail of the subject vocabulary. We find that supervised XMLC algorithms relying on transformer-based dense features give best results in terms of overall binary relevance metrics. However, focusing on graded relevance and performance in the long tail of our subject vocabulary, the LLM-based generative methods give better results, making them a promising alternative for future productive use.

In many operational time-series forecasting applications, such as crowd demand forecasting, the risk related to under-prediction is substantially higher than that of over-prediction. Accurate prediction of rare demand spikes plays a critical role in downstream tasks. Yet most time-series forecasters are trained with symmetric objectives (e.g., MSE, MAE) and evaluated primarily on aggregate error, which can mask failures in extreme-values and peak-timing predictions. We introduce Asymmetric Peak-Aware Loss (APAL), a simple, model-agnostic objective that (i) penalizes under-predictions more heavily and (ii) increases the training weight of peak regions within each forecast window. We further propose a peak-critical evaluation protocol that complements MAE/MSE with channel-wise tail error (Top-10% and Top-1%) and peak metrics (precision, recall, F1 under timing tolerance, and peak timing error). We evaluate APAL on long-horizon multivariate forecasting across five state-of-the-art backbones, with a focus on pedestrian demand forecasting using (i) a production-ready subset of the City of Melbourne pedestrian hourly count dataset and (ii) a beach visitor count dataset. The generality of the loss function for time-series forecasting is tested on additional benchmarks. Across peak-critical datasets and settings, APAL improves tail accuracy and peak-prediction quality while exposing a controllable trade-off with aggregate error, making it a practical solution when peak-prediction failures are the dominant operational concern.

LLM-based agents are increasingly deployed in multi-agent environments whose incentives can shape their behavior. We introduce The Energy Society, a minimal survival economy for studying how competitive and cooperative incentives affect emergent behavior when inference cost is directly tied to survival: Agents spend energy based on model size when generating tokens, regain energy by completing jobs or receiving donations, and deactivate if their energy reaches zero. We compare competitive and cooperative objectives against a baseline setting and several control variants. Across experiments, larger models consistently consume the most energy and spend more energy than they gain, even in those settings where token cost is not size-dependent. Cooperative incentives substantially alter behavior: agents donate to reactivate others, sometimes at the cost of their own survival, and job allocation changes. Ablations reveal that allowing agents to recommend actions to each other supports coordination and ambitious job selection, while memory helps agents calibrate risk from past outcomes. Agents rarely choose direct sabotage, but show more subtle signs of self-serving behavior in the competitive setting. The Energy Society is a compact testbed for studying the interaction between token costs and group incentives under a survival pressure. Source code is available at https://github.com/LucasBergholdt/EnergySociety

In future cities every driver may own a vehicle which could be either independently driven (HDV), or autonomously routed and piloted (CAV). The autonomous operations could be handled by a few competing companies. What is the market structure which would make this market aligned with city goals? In this paper we discuss a variant of the emerging market of collectively routed fleets of CAVs, where revenue for fleet operators is proportional to market share. We provide benchmark scenarios to compare the routing algorithms. We present several routing algorithms and demonstrate that, when the attitudes of human drivers towards CAVs exhibit significant diversity, randomised CAV routing, resulting in unpredictable travel times for HDVs, is more efficient than routing proportional to system optimum/user equilibrium. Based on this, we propose to improve the design of the market by augmenting the market-share objective with mean systemwide travel time in order to limit antisocial randomised strategies of fleet operators and drive the competition towards social welfare oriented cooperation.

Current voice AI benchmarks typically evaluate isolated capabilities such as speech intelligibility, word error rate, or text-based dialogue quality, but they rarely test whether systems harness the acoustic information that distinguishes spoken language from its textual representation. To this end, we introduce the Real World Voice EQ Bench, a multidimensional benchmark for evaluating voice AI across text-to-speech (TTS), speech-to-speech (STS), speech understanding (SU), and automatic speech recognition (ASR). Our evaluations indicate that performance is highly dimension-specific. For TTS, naturalness, expressiveness, identity stability, and reliability are largely independent evaluation dimensions. For STS, access to audio does not guarantee use of vocal affect, and some agents remain largely transcript-driven. For SU, models perform unevenly across paralinguistic tasks. For ASR, real world accent, emotion, noise, and conversational conditions expose failures that are not captured by established clean-speech benchmarks. Together, these results show that voice AI should be evaluated as a profile of acoustic, expressive, interactional, and robustness capabilities rather than by a single aggregate score.

We report on CoreForge, an experience in using large language models (LLMs) to build an unweighted MaxSAT solver from research papers rather than from an existing solver codebase. The project focuses on unsatisfiability-based MaxSAT algorithms and follows an iterative workflow that combines paper discussions with ChatGPT, implementation through Codex prompts, and repeated LLM-assisted code audits and revisions. Although the codebase implements several algorithms and solver components, our evaluation focuses on configurations that combine core-guided optimization, lightweight preprocessing, core minimization, integration with integer linear optimization backends, and a new core-sequence lookahead approach. Our experience suggests that LLMs can support solver implementation from papers, while requiring external validation, benchmarking, and human guidance. In our experiments, fuzzing and MaxSAT Evaluation instances did not reveal wrong answers in the tested configurations, although performance remains below the best hand-engineered MaxSAT solvers. We summarize what worked, what remained difficult, and the lessons for future LLM-assisted solver development.

Current evaluation of epistemic uncertainty relies on tasks such as out-ofdistribution detection and active learning. However, the Bayes-optimal decision strategies for these tasks do not coincide with the scores commonly used to quantify epistemic uncertainty. Building on the epistemic reject-option framework, we evaluate epistemic uncertainty using its ability to identify regret, the reducible error. Formulating selective prediction as a constrained optimization over coverage, expected risk, and regret, we prove the optimal selector is a thresholded convex combination of the ground-truth aleatoric and epistemic uncertainties. This theoretical unification exposes a weakness in recent uncertainty disentanglement literature: we demonstrate that standard correlation metrics between learned components do not necessarily predict their actual operational utility. We instead propose to evaluate the achievable risk, regret, coverage surface of the decomposition as a diagnostic for joint disentanglement and utility. Benchmarking standard methods on datasets with dense human annotations reveals that decision-theoretic rankings can disagree substantially with proxy-task rankings, including pairwise rank inversions between methods that are top-ranked on one criterion and bottom-ranked on other.

Large Language Models (LLMs) perform differently on identical programming tasks when prompted in different natural languages, a phenomenon known as language bias. While this behavior has been widely studied for general text generation, its impact on code generation quality and programming conventions remains largely unexplored. We investigate how the language used to describe programming tasks affects the source code generated by GPT-4o mini, DeepSeek, and Claude. Our study comprises 460 coding tasks spanning Python (230) and Java (230). We translate and manually curate the original English prompts into Chinese, Hindi, Spanish, and Italian while preserving their technical meaning. We evaluate the generated code using multiple dimensions, including functional correctness through test pass rates, structural quality using established code metrics, issues detected by static analysis tools, and lexical characteristics such as the language used in identifiers and comments. Our results show that (i) English prompts do not consistently produce the best functional correctness or code quality, (ii) the impact of prompt language depends on both the programming language and the LLM, and (iii) generated code frequently mixes English with the prompt language in comments and string literals. These findings provide the first curated multilingual benchmark for studying language bias in code generation and offer insights for developing more robust multilingual code generation systems.

Large language models are increasingly trained as interactive agents for long-horizon tasks involving multi-turn interaction, tool use, and environment feedback. Outcome-based reinforcement learning (RL) provides a practical optimization paradigm, but its sparse trajectory-level rewards offer limited guidance on intermediate decisions, leaving a supervision gap between episode-level outcomes and token-level policy learning. We propose SEED (SElf-Evolving On-Policy Distillation), a self-evolving framework that converts completed on-policy trajectories into training-time hindsight skills and distills their behavioral effect back into the policy model. SEED first fine-tunes the policy to analyze completed trajectories and generate natural-language skills that capture reusable workflows, decisive observations, or failure-avoidance rules. During RL, the current policy both collects trajectories and serves as the analyzer that extracts hindsight skills from them. Policy updates therefore improve subsequent decision making and skill analysis together, allowing hindsight supervision to evolve with the policy. SEED then re-scores the sampled actions under ordinary and skill-augmented contexts, converting the skill-induced probability shift into a dense token-level on-policy distillation signal. This signal is jointly optimized with outcome-based RL, keeping the auxiliary supervision aligned with the current trajectory distribution. Extensive experiments on text-based and vision-based agentic tasks show that SEED consistently improves performance and sample efficiency, exhibiting robust generalization to unseen scenarios. Our code is available at https://github.com/jinyangwu/SEED.

Knowledge graph question generation (KGQG) aims to generate natural-language questions from structured graph evidence. Existing KGQG benchmarks, however, are mostly built on static knowledge graphs and do not encode the temporal scopes of graph facts. As a result, they cannot evaluate whether generated questions faithfully preserve temporal validity, event ordering, and answer-determining temporal constraints. In this paper, we study temporal knowledge graph question generation (TKGQG), where a generated question must be faithful to both the support subgraph and the temporal constraints required to identify the target answer. We propose ChronoQG, the first temporally expressive and hop-bounded benchmark construction framework for TKGQG. ChronoQG integrates a comprehensive temporal-constraint taxonomy, topology-temporal subgraph sampling, and trace-grounded question generation to construct temporally faithful questions. The framework produces four benchmark datasets from heterogeneous temporal knowledge graphs, totaling 16,011 verified questions. We evaluate representative LLM-based KGQG methods and prompting baselines across diverse TKGQG settings, including temporal-constraint counts, topological templates, and temporal-constraint types. The results show that existing methods struggle to preserve temporal constraints, especially under multi-constraint settings and harder temporal-constraint types. These findings reveal a clear gap between static KGQG and TKGQG, and establish ChronoQG as a challenging testbed for temporally faithful question generation.

Vision-Language-Action (VLA) models have achieved impressive results in visuomotor policy learning, yet remain fundamentally reactive, mapping current observations and language to actions without explicit forward prediction of world dynamics. Existing visual foresight methods predict future visual states but lack explicit motion guidance: they show where to go but not how to get there. We argue that future feature prediction and sparse point tracking are naturally complementary: the former provides the goal state, while the latter captures the continuous motion path toward it. We propose FoMoVLA, a framework that augments VLA representations with explicit spatio-temporal supervision by jointly learning future feature foresight and sparse 2D point tracking, enhancing the continuous action policy. FoMoVLA introduces compact foresight tokens to decode future feature states, decodes sparse temporal 2D point trajectories to model compact geometric motion, and couples both through a lightweight future-conditioned cross-attention module that enables consistent reasoning between anticipated states and point dynamics. Extensive experiments on LIBERO, RoboCasa GR-1 Tabletop, and LIBERO-Plus demonstrate state-of-the-art performance and strong zero-shot generalization. Project page is available at https://liauto-research.github.io/FoMoVLA.

Transparent educational question answering asks for answers that are not only correct but explainable, and doing so with small models rules out the reasoning power of the largest proprietary systems. The EXACT 2026 competition poses this problem concretely: open-weight language models of at most 8B parameters, self-hosted, with a natural-language explanation for every answer. It pairs two tasks: logical reasoning over university regulations, and multi-step physics problem solving. We describe the system that team \cotu{} developed to address both, a neuro-symbolic Program-of-Thought pipeline in which a 4B backbone writes a program rather than stating an answer directly: for regulation queries it emits a Z3 encoding whose entailment verdict grounds the deduction, and for physics it emits numerical Python, both wrapped in a shared self-correction loop and a unified explained-JSON output. Answer-type routing, distillation-based task fine-tuning, and a latency-aware serving stack -- SGLang with speculative decoding -- keep the system within the 60-second per-query limit. The system achieved a \textbf{perfect score} on the physics task in both automated selection rounds and obtained the \textbf{highest final-round technical score} of any team -- $13.44/15$, combining automated answer evaluation with expert-judged reasoning depth -- with the equally weighted presentation score included, \cotu{} placed 3rd overall. Grounding answers in a symbolic solver yields correct, verifiable deductions at the 4B scale, and the residual difficulty lies in premise selection rather than the deduction itself.

Temporal Knowledge Graphs (TKGs) record how facts evolve over time, but forecasting future events on a TKG remains difficult for three reasons: (i) long-range temporal dependencies are hard to encode; (ii) events on different chains mutually excite or inhibit one another in ways that snapshot-level models cannot express; and (iii) inter-arrival times are heavy-tailed and statistically sparse, so deterministic time predictors are unreliable. We address these three issues with a single framework, the \textbf{Group Attention Neural Hawkes Process (GAttNHP)}, built around three matched components. First, a self-attention encoder casts each subject--relation chain as a continuous-time point process and captures the lingering excitation of distant history. Second, a semantic soft-grouping module turns globally learnable Hawkes priors into an analytical cross-attention mask, so chains share excitation patterns through their latent group memberships rather than through exhaustive pairwise computation. Third, a Non-Crossing Quantile (NCQ) regression head replaces mean-based time prediction, providing calibrated, monotonically ordered quantile estimates that remain stable under heavy-tailed inter-arrival distributions. On six benchmark TKG datasets, GAttNHP improves over state-of-the-art baselines on both entity prediction and time prediction, and ablations confirm that its largest gains arise on the long-tail event chains where existing models fail most severely.

Counterfactual explanations provide local, interpretable insight by identifying changes to an input that would alter its assigned outcome. Although well established in supervised learning, their extension to clustering is less direct, since cluster assignments are unlabeled and governed by the geometry of the partition. This paper introduces VoICE, a Voronoi-Induced Counterfactual Explainability framework for feature-weighted $k$-means clustering. Rather than treating cluster change as a crossing of a single pairwise centroid boundary, VoICE formulates counterfactual generation as projection onto the full weighted Voronoi region of a target cluster, incorporating feature weights directly into both the clustering geometry and the counterfactual objective to yield least-cost and parsimonious explanations under actionability constraints. Target regions are further intersected with data-derived bounds and homothetically contracted towards their centroids, limiting extrapolation and boundary sensitivity. VoICE consistently produces valid target-cluster membership, across several benchmark datasets, where the leading pairwise baseline does not.

Probably not, at least for meta-analyses in economics. In a pre-registered, identity-masked, within-paper experiment, the authors of 44 meta-analyses ranked three AI reports on their own paper by usefulness for improving it: a single pass by a frontier model against two multi-agent debate tools we built and expected to win. All reports were held to a common length and template. The authors preferred the single pass, by 0.66 rank points over mad-research (95% CI 0.32 to 1.00) and 0.57 over paper-workshop (0.16 to 0.95), though paper-workshop spent roughly thirty times the tokens. Authors who recalled their journal referee report usually placed it first and never last; in a separate exercise, three AI judges almost always placed the real journal referee report last. Among the three AI reports, Gemini (the judge whose model family wrote none of the reports) would have ranked paper-workshop first in the authors' place, reversing the single-pass preference. The reversal warns against substituting an AI judge for the author. We measure perceived usefulness for finished papers; whether AI should referee papers is a separate question.

We present for video understanding (classification) a split space-time attention model, VideoSEMA, consisting of a scalable and efficient Mamba-like attention (SEMA) block in space and a softmax temporal attention in time. In each frame, SEMA attention applies a local window attention in parallel with a global averaging in a Mamba macro-architecture, which is called Mamba-like. Under certain rank conditions, we prove that the computationally cheaper split space-time attention is equivalent to full space-time attention. On benchmark K400 data sets, VideoSEMA out-performs heavier vision transformer and Mamba models. On benchmark SSv2 data, VideoSEMA leads in top-1 accuracy among models of similar parameter sizes. As image resolution scales up from standard $224^2$ to $1024^2$ on K400 and without fine-tuning, VideoSEMA degrades much more gracefully than VideoMamba in accuracy. It is promising to extend VideoSEMA to longer videos with a dilated/sparse temporal attention.

Large language models routinely produce fluent answers to single-shot prompts, yet deploying them as reliable components of a domain decision system is substantially harder. Closing this gap is the work of harness engineering: the deliberate composition of deterministic scaffolding (symbolic filters, retrieval, schema-typed I/O, LLM-as-judge loops, HITL gates, persistent state, audit trails) around an LLM core. We present a case study in academic supervision, a domain combining high-stakes recommendation, longitudinal accountability, and structured operational workflows. We compare a baseline (ASA), a GPT-5 chatbot with no scaffolding, against a multi-module system (ASuS) that wraps the much smaller GPT-4o-mini in a LangGraph harness with symbolic-semantic retrieval, schema-validated outputs, LLM-as-judge with bounded retry, HITL gates, deterministic weighted risk scoring with LLM narration, and a per-node SQLite audit trail. The evaluation rubric is retargeted at six harness-mechanism dimensions (grounding, explainability, consistency, process integrity, cognitive load, constraint adherence). A blind ten-rater hybrid evaluation, supplemented by a 2 x 2 model-harness ablation, finds that ASuS, despite using a much smaller base model, outscores ASA on every dimension. Across ten raters the pooled mean for ASuS is 4.08 versus 1.23 for ASA, and 8 of 10 raters reject the null at alpha = 0.05 on a paired Wilcoxon test; full numbers are in Sections 6.4 and 6.7. The ablation confirms that the structural contributions of the harness are largely model-invariant. We extract seven recurring harness-engineering patterns and argue that where reliability, traceability, and institutional consistency matter more than open-ended fluency, harness engineering challenges the prevailing 'bigger model is better' intuition.

Multiple instance learning (MIL) has become the main paradigm for whole-slide image (WSI) analysis in computational pathology. However, existing MIL aggregators are still typically trained from scratch for each downstream task, relying on limited slide-level labels to learn both aggregation mechanisms and downstream discriminative representations simultaneously. As a result, they often suffer from unstable optimization, overfitting, and limited transferability. Similar to pretrained ResNet and Vision Transformer models in natural image learning, MIL also requires reusable pretrained initialization. However, high-quality slide-level pretraining data remain scarce, and MIL models are usually lightweight and weakly supervised, making large-scale pretraining difficult in practice. To address this challenge, we propose a distillation-based pretraining framework for MIL, which leverages two slide-level foundation models, TITAN and CARE, as teachers to transfer their representational knowledge into a diverse set of MIL architectures. To effectively balance supervision from different teachers, we further introduce an angular dispersion normalized distillation loss. The distilled weights are then used as initialization for downstream adaptation. We conduct systematic evaluations on 15 benchmark datasets under both linear probing and full-parameter fine-tuning, and further validate its advantages in few-shot scenarios. Experimental results show that pretraining generally improves MIL aggregators over from scratch training, especially in linear-probing and few-shot settings, while maintaining the computational efficiency of lightweight MIL models. Code is available at https://github.com/fu0201/MIL_Pretrained.

Understanding driver emotion and state is critical for the next generation of intelligent in-cabin systems that ensure safety and enhance human-vehicle interaction. However, existing public datasets for in-cabin affective computing are largely limited to visual modalities and rarely include conversational information, making it difficult to capture the linguistic and interactive cues underlying driver emotion. To address these gaps, we introduce InCarEmo, a multimodal dataset for in-cabin emotion recognition and driver state monitoring. InCarEmo integrates RGB and infrared video, in-cabin audio, and dialogue text collected from scripted in-cabin scenarios designed to simulate realistic driver behaviors, covering diverse lighting conditions and driving contexts. The dataset supports three primary tasks: 1) multimodal emotion recognition, 2) fatigue detection, and 3) distraction monitoring. In addition to the original Chinese data, we construct an auxiliary English benchmark to support preliminary cross-lingual evaluation. We provide a unified benchmark with extensive baseline results across unimodal and multimodal methods, including analyses under modality-missing and noise conditions. Experimental results demonstrate the benefits of multimodal fusion and reveal remaining challenges under real-world noise and low-light conditions. By releasing InCarEmo, we aim to establish a comprehensive foundation for robust, interpretable, and human-centric in-cabin affective understanding, promoting safer and more empathetic driver-vehicle interaction.

Efficient multimodal document question answering with explicit visual grounding, locating the precise document region that supports each answer remains an open challenge. Current approaches bifurcate into Supervised Fine-Tuning (SFT), which requires large annotated datasets and reaches optimization plateaus, and reasoning-centric Reinforcement Learning (RL), which depends on verbose intermediate traces that inflate inference token cost without clear benefit. We introduce Perception-RFT, a training framework that applies Group Relative Policy Optimization (GRPO) to multimodal document QA, bypassing intermediate reasoning tokens to directly align visual features with structured grounding outputs. To rigorously evaluate the necessity of reasoning, we construct a reasoning variant under identical reward settings. We find that reasoning-enabled models suppress their reasoning traces during training, converging to direct perception-based policies at the 4B parameter scale, reducing per-query inference token length by more than 60%, while reasoning-enabled RL underperforms perception-only training. Through a fine-grained analysis of Qwen3-VL-4B optimization dynamics, we confirm that SFT saturation and cold-start RL instability established in text-domain post-training extend to multimodal, and identify a previously uncharacterized Grounding Divergence: a selective trade-off between semantic robustness and geometric precision on two out of distribution (OOD) benchmarks (4,828 samples) under joint RL optimization. We further show that an early SFT$\rightarrow$RL transition achieves comparable precision with 65% less training data.

Evaluations (Evals) are a deployment bottleneck for real-world AI applications: public benchmarks rarely match a team's users, context, or policies, and human review is often tedious to scale. Motivated by our work with AI applications in the public sector, this project addresses recurring evaluation challenges encountered when applications must satisfy local policy and governance requirements. We present Kaleidoscope, an integrated workflow for contextual functional evaluation that links persona-based test generation, contextualized rubrics, and human review for reliability-gated automated scoring. Generated test cases are scored against application-specific rubrics; human annotations provide reviewable labels; and LLM judges automate scoring only when their agreement with those labels meets a configured threshold. Kaleidoscope is therefore a practical, inspectable, iterative workflow for product teams. We report early evidence from a three-week pilot across four organizational use cases and custom-rubric judge experiments on 108 annotated Q\&A pairs spanning four domains and 14 evaluation dimensions. The results highlight useful features for end-to-end reliable, automated scoring.

Continuous-time spiking neural networks (SNNs) provide an event-driven framework for temporal computation, computational neuroscience, and neuromorphic hardware. However, training deep continuous-time SNNs is severely constrained by the memory required for exact spike-time computation, which evaluates and retains candidate firing times over intervals determined by presynaptic spike ordering. Here we introduce a memory-efficient training framework based on differentiable spike-time discretization (DSTD) for leaky integrate-and-fire neurons with general membrane and synaptic time constants. DSTD maps irregular presynaptic spikes onto differentiable weighted events at fixed time points, replacing the input-dependent candidate dimension with $M$ fixed time intervals while accurately approximating continuous-time membrane-potential dynamics. This reduces candidate-related activation memory from $O(N_{\mathrm{out}}N_{\mathrm{in}})$ to $O(N_{\mathrm{out}}M)$ in the case of time-to-first-spike (TTFS) coding, where $N_{\mathrm{in}}$ and $N_{\mathrm{out}}$ denote the numbers of presynaptic and postsynaptic neurons, respectively. We further introduce synfire-chain-inspired temporal regularization that organizes layer-wise firing windows, mitigates dead-neuron failures, and enables pipeline-like processing. In dense LIF layers, DSTD reduced peak memory consumption by up to approximately 100-fold and training time by up to approximately 20-fold compared with exact spike-time computation. Together, these methods allowed us to train 9-layer convolutional SNNs on CIFAR-10 and 20-layer convolutional SNNs on Fashion-MNIST on a single GPU.

Deploying large language models (LLMs) as personal assistants on mobile devices demands privacy, low latency, and offline availability, yet the computational cost of giant models clashes with strict edge-hardware budgets. We argue that this tension cannot be resolved by model compression alone; it requires decomposing on-device intelligence into complementary functional roles. We present SmartRAG, a fully on-device framework that organizes an intelligent assistant around four coordinated modules -- Perception, Memory, Focus, and Thinking. At the core of SmartRAG is EvoNER, a continually learnable named-entity recognizer that incrementally expands its label inventory through teacher-distilled updates, enabling the system to absorb previously unseen entity types without retraining the backbone LLM. Extracted knowledge is stored in MRGraph, a three-layer provenance-preserving knowledge graph, and retrieved at query time through a hybrid pipeline combining graph traversal, lexical matching, and dense semantic search. The on-device LLM is invoked only for high-value semantic operations -- labeling, planning, and answer synthesis -- keeping inference costs bounded. Experiments on four QA benchmarks (TriviaQA, Natural Questions, HotpotQA, MultiHopQA) show that SmartRAG with a quantized 1.7B-parameter backbone achieves multi-hop reasoning performance competitive with models up to 18$\times$ larger, while running entirely on commodity smartphones within practical memory and latency envelopes.

While Multimodal Large Language Models (MLLMs) excel in general tasks, rigorous scientific reasoning remains challenging due to the limitations of monolithic, linear planning. Such sequential designs often suffer from visual-semantic misalignment, long-context hallucinations, and brittle execution under fixed task granularity. We propose TopoAgent, a self-evolving topological framework that replaces linear trajectories with dynamic, state-isolated graph evolution. TopoAgent first employs a front-end decomposer to fracture complex queries into visually-grounded atoms. These atoms are organized into a Directed Acyclic Graph (DAG) based on their dependencies, enabling strict context isolation to shield the reasoning engine from irrelevant historical noise. Furthermore, we introduce adaptive atomic fission, which dynamically splits bottleneck nodes into finer-grained sub-atoms at runtime when tool capability boundaries are exceeded. Extensive experiments across mathematics, physics, and chemistry benchmarks demonstrate that TopoAgent significantly outperforms state-of-the-art linear agent frameworks, providing a robust, noise-resistant, and self-correcting paradigm for autonomous scientific reasoning.

Persistent external memory enhances agent continuity but introduces persistent security vulnerabilities: adversarial content can be injected via standard interaction channels, retained across turns, and later distort downstream behavior. To address this challenge, we propose MemPoison, a comprehensive benchmark and analysis framework featuring 1227 hand-validated cases across four attack types, three injection channels, and three representative memory substrates, evaluated on seven open-weight and three closed-weight model families. We introduce a three-tier taxonomy: (L1) direct single-record corruption, (L2) compositional multi-record corruption and (L3) context-triggered dormant corruption. Our evaluations reveal a distinct defense frontier: while baseline write-time defenses, such as consistency checks, substantially suppress direct L1 attacks, they fail to reliably suppress L2 and L3 attacks. Through mechanistic influence decomposition (MID), we demonstrate structural blind spots in write-time defenses, which admit seemingly benign records that later become harmful through joint retrieval composition or trigger-conditioned activation. Our findings advocate for shifting from static filtering to adaptive, context-sensitive memory defense strategies.

As Model Context Protocol (MCP) servers emerge as the core infrastructure for connecting LLMs with external tools, existing benchmarks leverage real-world MCP servers to evaluate LLM agents' tool-using capabilities. However, these benchmarks overlook the continuous evolution of tool interfaces and functionalities within MCP servers, resulting in flawed assessments that fail to capture the agent's adaptability in changing tool landscapes. To bridge this gap, we introduce \textbf{MCPEvol-Bench}, a novel benchmark for evaluating the task-solving capabilities of LLM agents under dynamic toolset evolution. Inspired by large-scale empirical study, we propose 11 mutation operators to simulate realistic tool evolution within 123 MCP servers. We benchmark 12 state-of-the-art LLMs on multiple versions of MCP servers, revealing that even frontier models struggle to adapt to evolving tools. For instance, GPT-5.4 and Claude-Sonnet-4-6 exhibit performance declines of 13.7\% and 14.4\% in evolved MCP servers, respectively, accompanied by substantial increases in planning and reasoning errors. These findings highlight the vulnerability of LLM-driven workflows, establishing MCPEvol-Bench as a standard for evaluating agent adaptability in dynamic tool environments.

Action supervision in vision-language-action (VLA) models is often treated as a downstream objective for learning action prediction. In this paper, we study it instead as a force that shapes inherited multimodal representations. We show that this shaping has a dual effect: it is necessary for forming action-compatible representations, but when action supervision is applied too directly to the inherited multimodal pathway, it can also destabilize representations that support language-side processing and object grounding. To address this tension, we introduce Action QFormer, a query-based action-facing interface that uses instruction-conditioned queries to reorganize inherited multimodal information into action-facing representations before downstream action generation. In zero-shot sim-to-real navigation, Action QFormer improves average closed-loop task success from 18.8% to 56.3%, raises fixed-instruction action-generation correctness from 22.5% to 75.5%, and nearly eliminates out-of-distribution instruction generations. Further analyses show that Action QFormer changes how action supervision shapes inherited multimodal representations, reducing broad upstream rewriting while preserving targeted and sometimes constructive action-supervised adaptation. These results suggest that improving VLA performance requires not only stronger pretrained backbones, but also better ways of selecting and organizing inherited multimodal information while controlling how it is shaped under action supervision.

Inferring apparent personality from facial images is important in social scenarios for embodied agents in human-robot interaction. Unlike inferring intrinsic personality traits via conversation, this task models first-impression personality perception based solely on facial appearance before interaction begins. Existing studies mainly focus on the Big Five personality model and often rely on language or multimodal inputs. As a result, it remains unclear whether facial cues alone can support meaningful associations with perceived personality traits. This question is particularly relevant for MBTI types, which are widely used in practice and more readily interpretable by large language models. To this end, we propose \textbf{GlanceFace}, an end-to-end framework for apparent personality inference leveraging vision-language models to introduce semantic priors and a semantic-enhanced facial representation module to capture subtle personality-related cues, together with an uncertainty-aware learning strategy to handle noisy and subjective annotations. Extensive experiments demonstrate strong performance on MBTI-based apparent personality benchmarks and reveal relationships between facial characteristics and perceived personality traits, highlighting its potential to support adaptive initial interaction strategies for embodied agents. The code and dataset are available at https://github.com/MrHuan3/GlanceFace.

Low-bit GEMM is increasingly central to efficient ML inference, yet very-low-bit execution remains a poor fit for conventional CPUs. Practical deployment spans fragmented regimes-from 1/2/4-bit weights to varying activation precision-whose feasibility, reuse opportunity, and support cost differ under fixed SIMD and register-file budgets, making lightweight CPU support selection a first-class design problem. We present ExaGEMM, a workload-aware codesign and exploration framework for CPU-native low-bit GEMM via register-resident LUT execution. The key insight is that existing SIMD datapaths already cover table generation and accumulation; the only new hardware is an in-register select/feed mechanism with explicitly modeled cost. ExaGEMM co-explores parameterized kernels and lightweight SIMD ISA support using analytical models of register feasibility, compute cost, memory traffic, and hardware overhead, pruning the candidate space by 99.2% before simulation. It then identifies non-dominated support points and generates ISA specs, gem5 patches, and GEMM kernels for validation. Across representative ML models and CPU targets, ExaGEMM improves latency by 13.29x over software-only baselines, while showing that workload-aware frontier selection is especially important for mixed-precision LLM workloads.

Vision--language models have become increasingly capable of interpreting visual scenes, but it remains unclear whether they can use information to make strategically effective decisions. We investigate this question in soccer, where models observe the seconds preceding an on-ball decision and must choose whether to shoot or pass to a specific teammate. Unlike conventional visual-understanding tasks, soccer enables decisions to be evaluated quantitatively by estimating the value of every available action. We introduce SportD, a benchmark comprising 478 on-ball decisions from the 2022 FIFA World Cup. Each model choice is evaluated against a possession-value model that estimates the action that most increases the attacking team's probability of scoring, allowing us to measure both optimal-action accuracy and the value forfeited by suboptimal decisions. Across three frontier VLMs, the best selects the highest-valued action on 31.4% of events, compared with 38.9% for the professional players, and all models incur significantly greater regret. Further analysis reveals a systematic preference for lower-variance and lower-reward actions: VLMs shoot less often and select substantially less progressive passes than either the optimal policy or the real players. The models also reproduce the player's specific action above chance even when that action is suboptimal, suggesting partial imitation of familiar play patterns rather than consistent evaluation of counterfactual alternatives. SportD provides a value-grounded testbed for measuring physical strategic reasoning in VLMs.

Reinforcement learning with verifiable rewards (RLVR) commonly uses entropy for advantage shaping. However, entropy cannot distinguish useful uncertainty from detrimental confusion, limiting its effectiveness as a correctness signal. We propose Contrastive Policy Optimization (CPO), which uses token-level contrastive disagreement between reference-guided and vanilla generation distributions for correctness-aware advantage shaping. Both theoretical and empirical results show that this disagreement reliably indicates token-level correctness. We further show that On-policy Distillation is a special case of CPO, where the posterior distribution is instantiated by an external teacher model. CPO also resolves the zero-advantage problem. Experiments on in-domain and out-of-domain benchmarks demonstrate that CPO substantially outperforms entropy-based RLVR methods while maintaining strong generalization. Further analysis shows that correct and incorrect responses naturally support exploration and exploitation respectively, and balancing both leads to the best performance.

Long-tailed label distributions reduce the reliability of deep learning for electrocardiogram (ECG) arrhythmia diagnosis, particularly for clinically important but rare abnormalities. Existing rebalancing and logit adjustment methods mainly address class frequency while overlooking direction-dependent morphological variability across ECG classes. This study proposes Angular Gaussian Supervised Contrastive Learning (AG-SCL) for long-tailed multi-label ECG diagnosis. AG-SCL integrates three components into a unified framework: an Angular Gaussian contrastive branch that models full-covariance class uncertainty on unit-normalized embeddings, Adaptive Logit Adjustment that learns bounded label-state-specific prior corrections instead of fixed frequency-based margins, and tail-aware augmentation that generates morphology-preserving views while protecting the 7-25 Hz QRS-dominant band. The method was evaluated on the public PTB-XL benchmark and a nocturnal ECG dataset comprising 1317 hours of recordings from 141 subjects. AG-SCL achieved the best macro-level performance on both datasets. On PTB-XL, it obtained a balanced accuracy of 0.838, sensitivity of 0.709, specificity of 0.968, mean average precision of 0.495, and TPR at 5% FPR of 0.778. On Noc-ECG, the corresponding values were 0.918, 0.889, 0.947, 0.488, and 0.900. The largest gains occurred in rare or morphologically unstable rhythm classes, while ablation studies confirmed the contributions of full-covariance modelling, Adaptive Logit Adjustment, and tail-aware augmentation. AG-SCL improves long-tailed ECG diagnosis by combining prior calibration with anisotropic representation learning, enhancing sensitivity to rare arrhythmias while maintaining clinically relevant specificity. Our code is available at: https://github.com/Open-EXG/AG-SCL-for-Long-Tailed-ECG.

A growing class of agentic systems maintain persistent state across sessions through memory files, behavioral preferences, and knowledge bases. While this makes agents more useful and self-improving, it also creates a new attack surface for prompt injections in which malicious instructions can be embedded within persistent files and influence future behavior. In this work, we study prompt injection attacks in memory-based agentic systems using a sandboxed synthetic workspace. We evaluate two agentic systems, Anthropic Claude Code and OpenAI Codex, across four models: Claude Haiku 4.5, Claude Opus 4.7, GPT-5.2, and GPT-5.5. Our results show that although it is difficult to make an agent overwrite its own memory files using untrusted external content, payloads already planted in those files can successfully attack current and future sessions. Attack success and payload persistence vary substantially across systems, models, adversarial goals, and multi-session attack sequences. These findings show that persistent memory changes the threat model for prompt injection and motivate defenses that protect memory updates without removing useful agent adaptation.

As conversational AI systems are designed for repeated use, a central question is how a series of interactions becomes a relationship. We present a longitudinal multimodal study of a memory-augmented conversational agent (24 participants x 10 sessions), in which participants rated five relational constructs -- familiarity, self-disclosure, perceived memory, conversational quality, and enjoyment -- after each session. Two complementary dynamics emerge. First, conversational quality strongly shapes how enjoyable a session feels in the moment but does not carry forward across sessions, whereas perceived memory is relationally conditioned -- predicted by prior relational state rather than reflecting system capability alone -- and it shapes later enjoyment indirectly, via subsequent self-disclosure. Second, relationships are punctuated by discrete turning points -- crashes and surges -- that are partially traceable in multimodal behavior and open different intervention windows: surges are more behaviorally detectable in the moment, enjoyment surges persist more reliably than enjoyment crashes recover, and some crashes are better forecast from person-specific behavioral drift than detected after they have already occurred. Together, the findings suggest that longitudinal human-AI relationships are built through both slow accumulation and abrupt turning points.

This study examines feedback in English as a Foreign Language (EFL) writing contexts, focusing on written corrective feedback (WCF). Large language models (LLMs) can provide WCF at scale, but aligning them with pedagogical best practices remains an ongoing challenge. WCF meeting criteria like factuality or relevance may still be unsuitable for learning contexts, highlighting the need for extrinsic evaluation based on the learner's perspective. We deployed WCF systems in a university-level EFL class with nearly 2,000 students, collecting over 20,000 drafts. We evaluated the generated WCF from two perspectives: intrinsic evaluation by experienced English teachers using a rubric, and extrinsic evaluation via student feedback and engagement metrics. Results revealed low alignment between teacher expert ratings and student feedback. These findings suggest that traditional expert evaluation alone may not fully capture WCF's usability or helpfulness from the learner's perspective, highlighting the importance of learner-centered evaluation frameworks for AI-based applications in language education.

Existing LLM-based theorem provers have achieved impressive results on formal mathematics benchmarks, yet they remain confined to acting as autonomous agents that prove a stated proposition. In this paper, we propose MathCoPilot, a human-in-the-loop system that embodies a new human--AI symbiotic paradigm for mathematical research, in which the mathematician steers the high-level mathematical direction while AI agents carry out the detailed formalization and proof work under continuous human guidance. MathCoPilot unifies three core capabilities: (1) an interactive workbench where the mathematician and AI agents collaborate through a living proof blueprint that decomposes a proof into navigable steps the human can directly inspect, direct, and refine; (2) automated proving skill orchestration with adaptive knowledge base search and Lean-integrated iterative verification; and (3) topic-driven paper retrieval and automated formalization into a verified Lean knowledge base. Using MathCoPilot, we systematically compare four state-of-the-art LLMs, including Gemini~3.1~Pro, GPT-5.4, and Claude~Opus~4.7, on a FormalMATH subset and on two real PDE theorems requiring deep domain expertise, evaluating their ability to produce verified Lean~4 proofs and to identify errors in deliberately incorrect proofs. Our results show that while current models can handle undergraduate-level problems with high success rates under favorable autoformalization conditions, substantial challenges remain for domain-specific theorems requiring genuine mathematical understanding.

Collective problem solving often requires that group members consider the tradeoff between exploitation of known solutions and exploration for new ones, where information of known solutions can be disseminated among individual members through communication networks. The Mason--Watts experiment (PNAS 2012) showed that human groups in shorter-path networks outperform those in longer-path networks on a two-dimensional search task. In this work, we focus on the investigation of such network-efficiency effects in the setting of a group of large language model (LLM) agents. Specifically, we consider groups of sixteen LLM agents playing the Mason--Watts experiment on the eight Mason--Watts network topologies. Moreover, we develop mechanistic Bayesian optimization agents such that the performance of LLM agents can be compared with both the mechanistic agents and the human experimental data. Our computational experiments indicate that the LLM agents show a significant network-efficiency effect when instructed to randomize their first-round choices, but not under the default initialization. In this experiment, adding a one-sentence first-round randomization instruction improves collective payoff by more than three times the estimated payoff difference across the eight network topologies. Also, the Bayesian optimization agents obtain higher payoffs than the evaluated LLM agents on this spatial search task. We further compare the agents' exploration--exploitation behavior, copying, and spatial diversity.

Payment integration is a demanding repository-level software task: agents must select a suitable product, implement coordinated client-server flows, verify payment outcomes, and preserve consistency between transaction and business states. We introduce Alipay-PIBench, a benchmark for evaluating coding agents on realistic Alipay payment integration. It contains nine product-specific projects and 18 task instances, each organized into Basic functional-completion and Advanced risk-aware hardening scenarios. Scenario-specific rubrics support deterministic static, unit, integration, and end-to-end checks, supplemented by LLM-assisted assessment for semantic requirements. We evaluate six coding-agent models and report rubric pass rate (RPR). Under the with-skill condition, mean RPR ranges from 68.58% to 91.37%. Access to the alipay-payment-integration skill improves mean RPR by 10.31 percentage points on average relative to the without-skill condition, with gains varying across models, products, and scenarios. Method-level results distinguish source-level completion, executable payment behavior, and payment-domain requirements. Alipay-PIBench provides a controlled setting for diagnosing model capability and evaluating structured guidance in payment integration.

AI software development agents are increasingly capable of modifying infrastructure and security critical systems, creating risks where an agent completes its assigned task while covertly weakening safeguards through actions such as broadening permissions, degrading logging, or introducing persistence mechanisms. While frontier laboratories may deploy sophisticated monitoring pipelines, many organizations and individual users adopting coding agents lack the resources and governance maturity required to maintain complex learned monitor ensembles. We study this problem as infrastructure level sabotage under task success in the infrastructure-as-code setting from ControlArena and compare trained and untrained monitors across synchronous and asynchronous deployment modes. We introduce an Information Flow Graph (IFG) monitor that analyzes structural security regressions using control-flow and data-flow graph diffs alongside raw code diffs. In asynchronous evaluation, an untrained git diff monitor misses 11.6% attacks at 1% false positive rate auditing budget, our untrained IFG monitor reduces this to 3.5%, and the trained Async RF monitor achieves below 1%. Unlike the trained Async RF monitor, IFG can also operate synchronously as a pre-deployment safeguard that blocks suspicious infrastructure changes before execution or merge. In synchronous mode, IFG realtime rollback reduces joint main task and covert task success from 74.4% to 0.0% with no measurable reduction in legitimate task completion. These results suggest that untrained structural monitors provide a practical and auditable path toward democratizing deployment safety for organizations seeking trustworthy adoption of advanced AI agents.

A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.24031). This report keeps the hardware and asks what a model that fits can do: we deploy MiniCPM-V-4.6, a modern multimodal assistant pairing a SigLIP2 vision encoder and window-attention merger (16x visual token compression) with a compact hybrid gated-delta-net backbone, entirely on the GPU. Three results. (i) An all-GPU engine built on measured foundations: projections that dequantize 8-bit weights once and call the vendor SGEMM still in the last Fermi toolchain (64% of FP32 peak; our best hand-written GEMM hit 37%, wrongly called the ceiling); a chunked delta-rule rewrite of the recurrent layers, 2.8x faster than the sequential scan once attribution exposed one bad kernel; and a measured negative: 4-bit weights make decode slower than 8-bit here, since Fermi issues nibble-unpacking shifts at half rate. (ii) The vision side is a port with a proof obligation: we translate tower, merger, and projector to sm_20 CUDA, validating every stage against a locally generated reference forward (full tower 1.4e-5). One failure, position-embedding bucketization differing on exact rational ties, generalizes to a rule: float tie-breaking in index arithmetic is implementation-defined; call the reference operator, do not reimplement it. (iii) Long context exposes an O(N^2) wall short benchmarks hide: prefill falls from 114 tok/s at 2k tokens to 21 at 10k in a naive attention kernel; per-head vendor-GEMM calls writing into the existing score buffer (zero extra memory) restore a flat profile (408 at 2k, 361 at 10k; 17x), verified by exact needle retrieval from 60% depth. The same rewrite cuts image encoding 6x, to 0.93s. The system answers an image question end-to-end in 1.7s.

Large language models (LLMs) have demonstrated strong reasoning performance, but their tendency to hallucinate limits their reliability in knowledge-intensive tasks requiring up-to-date and grounded information. Combining knowledge graphs (KGs) with LLMs facilitates the use of explicit symbolic knowledge that can be continuously updated without costly fine-tuning, while benefiting from rapidly advancing LLM reasoning. We propose MARS, a scalable knowledge graph question answering (KGQA) approach that requires no model fine-tuning. Rather than relying on open-ended agentic exploration, MARS performs a structured retrieval procedure that links question entities to the KG and iteratively retrieves relevant next-hop information. At each step, MARS decides whether to continue graph traversal or to generate the final SPARQL query, allowing the model to adapt the retrieval depth to the question while keeping the overall pipeline more predictable than fully agentic approaches. We evaluate MARS on three established KGQA benchmarks across several LLMs and settings, including multilingual evaluation, and provide insights through ablation studies and error analysis. Our approach achieves competitive performance relative to state-of-the-art methods while remaining efficient and scalable. The evaluation results, code and resources are publicly available: https://github.com/dice-group/mars-kgqa.

Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, yet their inference efficiency is significantly hindered by fixed-length generation constraints. Since the actual output length is unknown, output sequences are padded to a predefined maximum length, resulting in substantial redundant computation over unnecessary [EOS] tokens. In this work, we discover that DMLLMs implicitly reveal their valid semantic boundary at the very first denoising step through a distinct shift in MLP activation sparsity. Leveraging this observation, we propose Seer, a training-free framework that detects this boundary using a Signal-to-Noise Ratio (SNR)-based criterion and performs one-shot truncation of the redundant suffix for all subsequent computations. To preserve these theoretical gains during batched serving, Seer incorporates a hybrid execution strategy that maximizes throughput while seamlessly accommodating dynamic sequence lengths. Experimental results demonstrate that Seer effectively eliminates padding waste, accelerating throughput by up to $\sim$31$\times$. Across 9 benchmarks, Seer robustly maintains overall performance and even improves accuracy on complex visual tasks by mitigating noise leakage (e.g., DocVQA score increases from 63.52 to 63.66), offering a highly efficient, plug-and-play solution for DMLLM acceleration.

Modern industrial environments increasingly run many autonomous subsystems at once - schedulers, energy managers, vehicle fleets - each pursuing its own goals while sharing the same physical resources. Because high-level human intentions are translated into low-level control logic and then discarded, no running component can tell whether it is still doing what was actually intended, and goal conflicts surface only after they have caused a missed target or a shutdown. We propose the Intention Abstraction Layer (IAL), a domainagnostic middleware that represents intentions as first-class, persistent, and explainable runtime objects: a large language model grounded in a formal OWL ontology parses naturallanguage goals into structured intentions, a consistency monitor detects conflicts at registration time, before execution, and a transparency module explains them in natural language. We report a first proof of concept in which two autonomous agents register conflicting production and energy intentions, and the IAL flags and explains the conflict before it reaches the execution layer. The result is a mechanism that shifts behavioral assurance for cooperating autonomous systems from post-hoc failure analysis to pre-execution, intention-level checking.

Logistics systems increasingly mix \emph{autonomous logistic equipment} (ALE) with non-autonomous machinery under a central control system (CS), where the best decision-maker depends on who holds the most current world model, yet authority is fixed at design time. When an ALE's local model and the CS global model diverge, both act on incompatible beliefs and produce deadlocks that resource-based handling neither explains nor prevents. We propose the World-Model-Aware Responsibility Framework (WMARF), which assigns authority dynamically from CS world-model quality and equipment automation level, and classifies deadlocks by the state of authority -- none, in transition, or divergent. In a discrete-event simulation of two ALE converging on a semi-automated transfer point, reproduced over the VDA~5050 interface, a divergence deadlock under static control is prevented by a proximity-triggered handoff. Because authority follows information quality rather than a shared protocol, the scheme stays valid as autonomy grows.

Hyper-Connections (HC) expand the residual stream of Transformers into $N$ parallel streams, providing a form of memory scaling beyond model width and depth. Manifold-Constrained HC (mHC) stabilizes this formulation at scale. The large gains from $N{=}1$ to $N{=}4$ suggest residual-stream expansion as a promising scaling axis. However, existing HC-family methods typically stop at $N{=}4$. Our experiments reveal why: scaling mHC beyond this point yields diminishing performance gains and rapidly increasing training cost. We attribute this limitation to two bottlenecks: insufficient write-back information for an expanding number of streams and residual-mixing generation whose cost scales cubically with $N$. To address both bottlenecks, we propose xHC (Expanded Hyper-Connections), the first HC-family method to achieve meaningful expansion beyond $N{=}4$. xHC combines temporal feature augmentation for richer write-back with a sparse residual-stream architecture that updates only $k=4$ of the $N=16$ streams while retaining dense access to the full residual state. Across 18B and 28B MoE models, xHC delivers strong and consistent downstream improvements. On an 18B MoE model, xHC improves the average downstream score by 4.0 points over mHC, while adding only modest training FLOPs over the vanilla baseline. Scaling-law experiments show that the vanilla and mHC require $1.50\times$ and $1.19\times$ the compute of xHC, respectively, to reach the same loss. Practical large-$N$ training also requires controlling memory traffic from the expanded residual state. We therefore introduce xHC-Flash, which reduces the per-sublayer memory traffic from $73.5C$ to $40C$, comparable to the $34C$ required by mHC at $N{=}4$, while retaining the gains of full xHC. Together, xHC and xHC-Flash make large-$N$ residual-stream expansion effective and practical for LLM pre-training.

Large language models (LLMs) frequently contradict themselves when the surface form of a logically equivalent question changes. We present a benchmark of 350 question families (1,750 total questions) for Controlled Reformulation Testing (CRTBench) to evaluate logical invariance. In this benchmark, we investigate LLMs' ability to maintain consistent answers across controlled reformulations, which include contrapositive rewriting, double negation, negation flipping, and passive voice. We evaluate several frontier LLMs and observe an accuracy-consistency gap where GPT-5.4-mini achieves $98.9\%$ base accuracy but only $60.3\%$ family-level consistency, while reasoning-optimized o4-mini achieves $96.9\%$ consistency. From our experiments, we observe that failures cluster around logically nontrivial transformations such as contrapositive rewriting ($72.4\%$ for GPT-5.4-mini) and double negation ($84.6\%$), while surface-level rephrasing remains robust ($94-100\%$). Increasing reasoning effort improves GPT-5.4-mini to $85.4\%$ consistency, but leaves GPT-5.4 unchanged overall because gains on nested negation are offset by failures on quantifier families. These results show that accuracy alone is not enough for evaluating logical reasoning in LLMs.

This study presents WrAFT, a Writing Assessment and Feedback Tool, that delivers both accurate and reliable scores and effective comprehensive feedback to argumentative essays. WrAFT adopts a modular design by dividing automated writing evaluation (AWE) tasks into scoring, surface-level feedback, and deep-level feedback. In building the system, various Large Language Models (LLMs) have been evaluated, including LLaMA-3.3-70B-Instruct, GPT-4o, and Claude 3.7, through both direct prompting and supervised fine-tuning approaches. A proprietary dataset of 480 TOEFL Independent Writing essays with official benchmark scores was utilized. Benchmark-based evaluation shows that WrAFT achieves state-of-the-art performance in scoring, with a quadratic weighted kappa (QWK) of 0.84 and a root mean square error (RMSE) of 0.44 against official scores on a scale of 0-5. Human evaluation of system-generated feedback also reveals high approval ratings: 96.14 percent for surface-level feedback, 93.03 percent for deep-level macro feedback, and 94.69 percent for deep-level micro feedback. An interactive user interface has been developed for the system and is publicly available and free to use.

We formulate reinforcement learning (RL) in continuous time with discrete state spaces and possibly arbitrary action spaces via a stochastic control approach, where the state dynamics are modeled as a controlled continuous-time Markov chain (CTMC). We consider policy optimization problems and derive the corresponding policy gradient methods, leading to continuous-time variants of proximal policy optimization (PPO) and group relative policy optimization (GRPO). As a primary application, we develop a complete continuous-time RL framework for fine-tuning score-based discrete diffusion models. The proposed framework enables reward-driven optimization without requiring differentiability on the reward signals. In contrast to the existing GRPO-based approaches that only rely on terminal rewards, our formulation allows intermediate reward or advantage signals to be incorporated throughout the denoising trajectory. Importantly, when specialized to masked diffusion models (MDMs), our framework encompasses a rich class of policy parameterizations over the vocabulary simplex with analytically tractable probability ratios, providing a unified perspective on exploration and policy optimization in MDMs. For masked diffusion large language models (dLLMs), we further propose trajectory subsampling techniques to efficiently estimate computationally prohibitive trajectory likelihoods, reducing the computational cost of computing per-position probability ratios. We showcase the effectiveness of our methods on both low-dimensional entropy-regularized optimization problems and RL post-training of dLLMs on mathematical reasoning and coding tasks.

Multi-step retrosynthesis planning seeks to decompose a target molecule into commercially available building blocks through a sequence of feasible reactions. The vast combinatorial search space makes this task challenging even for expert chemists. Traditional methods combine tree search with offline-trained value networks that score candidates in isolation, without reasoning about complete multi-step routes. Recent work leverages Large Language Models (LLMs) for this task, but relies on simple interfaces that limit exploration of the full search space. We introduce RetroAgent, an LLM agent that bridges symbolic search and neural reasoning through a harness with structured memory. Through memory and chemistry tools, the agent observes the full search state, including explored routes, available alternatives, and properties of intermediates, enabling informed decisions grounded in both global progress and domain knowledge. Experiments on in-distribution and out-of-distribution benchmarks demonstrate that RetroAgent delivers strong performance and generalization.

This paper describes DS@GT ARC's third-place solution to the PlantCLEF 2026 challenge on multi-species plant identification in vegetation quadrat images, where systems must predict every species present in high-resolution (~3000 x 3000 pixel) plot photographs while training only on single-label images of individual plants. The pipeline is built around a fine-tuned DINOv2 ViT-L/14 classifier applied over a multi-scale tile decomposition of each quadrat, with per-tile predictions blended with a FAISS kNN retriever and post-processed by source-aware temporal fusion across repeated plot visits, a habitat-fit demotion that injects geographic and altitude priors from the training data, and a South-Western Europe geographic mask. Habitat-fit demotion and multi-scale aggregation are the largest individual contributors in the ablations. Two complementary training-centric directions, a cross-region transformer with noisy-student distillation on the LUCAS dataset and a label-as-query transformer decoder over synthetic CLS-domain pseudo-quadrats, yielded null results. An inference-time augmentation with instance-aware segmentation crops also did not improve performance. The selected submission reaches a private-leaderboard macro-F1 of 0.43902 (third place; public 0.51096); an unselected configuration of the same pipeline scored above 0.45 on the private set. Code: https://github.com/dsgt-arc/plantclef-2026.

Complex knowledge base question answering (KBQA) is commonly approached through either information retrieval over a question-specific subgraph or semantic parsing into an executable logical form. We study the latter paradigm. Recent large language model agents make semantic parsing interactive: they alternate between reasoning, querying the knowledge base, and extending a partial SPARQL query. This interleaving reduces reliance on one-shot generation, but makes the quality of \emph{KB grounding} depend on what the interaction tools expose. Existing agents retrieve or prune candidate properties mainly through lexical relevance and instance-level observations, without systematically conditioning on entity types, property domains and ranges, or the expected answer type. We call this failure mode \emph{type-blind grounding}. It enlarges the grounding search space and often produces plausible-looking but semantically incompatible triple patterns that execute to empty results. We propose SAGA (\underline{S}chema-\underline{A}ware \underline{G}rounding for \underline{A}gentic Text-to-SPARQL Generation), a training-free framework that turns property exploration into a schema-constrained grounding operation. SAGA maintains a persistent bidirectional type state, filters known-incompatible property candidates at construction time, presents the remaining graph patterns in a compact schema-annotated format, and handles missing schema information permissively through empirical and trace-local evidence. Across nine benchmark settings over Wikidata and Freebase, SAGA achieves the highest F1 on all nine settings and the highest exact-match accuracy on eight, while reducing empty-result queries across all reported Wikidata settings.

LLM evaluators (trained reward models and prompted LLM-as-a-Judge) are routinely validated via pairwise accuracy. In a multilingual setting, this operates under the premise that high pairwise accuracy implies reliable, language-neutral scoring. We show that this assumption does not hold. We conduct experiments with semantically identical instruction-response pairs across 23 languages, and find that multilingual evaluators assign significantly different scores to different evaluation languages. The bias is statistically significant and consistent across eight open-weight evaluators of different architectures and training paradigms, persists in frontier judges, and is strongly correlated with language resource level: lower-resource languages are scored more generously. Meanwhile, these biases are invisible to pairwise accuracy: evaluators achieve above 90% pairwise accuracy, yet have up to 43% difference in acceptance rate across languages under a global decision threshold, meaning, for instance, that harmful content in lower-resource languages is more likely to pass safety filters. Per-language thresholds would require language identification, which can be defeated by code-switched prompts. We then investigate why lower-resource languages receive higher rather than lower scores, and we find that model uncertainty is linked with the effect: models tend to give higher scores when less confident, both under negative log-likelihood and under token-free uncertainty measures; however, language identity remains a significant predictor after controlling for uncertainty, and the bias cannot be explained away by content difficulty alone, but is a structural, language-level misalignment.

This paper details the DS@GT ARC team's approach to BirdCLEF+ 2026, multi-label detection of animal vocalizations in soundscapes from the Pantanal wetlands. The 2026 edition adds about an hour of labeled soundscapes, shifting the task toward supervised pipelines fit to the labeled set. First, we build a competitive supervised baseline that ensembles a frozen Perch v2 backbone, a trained HGNetV2-B0 sound-event-detection network, and a non-bird prototypical head, reaching a private leaderboard score of 0.936 at rank 1894 within a 90-minute CPU budget. Second, we ask whether token-based representations can compete, contrasting codec representations from neural audio codecs against semantic representations from foundational embeddings. We compare two bioacoustic specialist models against four token-based encoders trained on AudioSet. The repository for this work can be found at https://github.com/dsgt-arc/birdclef-2026.

Robots are increasingly integrated into everyday contexts, including museums, where they can both entertain and educate visitors. To enhance visitor experience and engagement, we present a novel mixed-agent tour guide system that combines a physical robot with a projected virtual agent that actively participates in the tour through conversation and interaction, achieving the interaction richness of two mobile agents from a single platform. We validate the system through a within-subjects study with 30 participants to assess engagement, quality of experience, and learning performance. Participants experienced different conversational styles and agent configurations, and data were collected via surveys, behavioral sensors, and interviews. Results showed that engagement and quality of experience remained consistent across conditions. Learning performance revealed a significant gender-moderated difference: the mixed-agent conditions improved learning performance for female participants. This suggests that the proposed dyadic conversational style in this paper influenced learning performance differently by gender. Nonetheless, in interviews, participants reported a greater preference for mixed-agent teams regardless of gender, citing interaction as a key factor in their experience.

Generalist robot manipulation policies trained on large, diverse datasets have shown remarkable promise across a wide range of tasks. However, rigorously evaluating these policies remains a fundamental challenge. Real-world performance depends on a large combinatorial space of task factors including object poses and camera viewpoints, making full, exhaustive evaluation intractable. Additionally, real hardware evaluation is slow and resource-intensive, so current practice is to use narrow test suites that can miss critical failure modes and misrepresent true deployment readiness. We propose an active evaluation framework that addresses this challenge by treating policy evaluation as a sequential experimental design problem. Our approach fits a probabilistic surrogate model over a structured space of task factors and adaptively selects evaluation configurations to maximize information gain over the policy's performance distribution, allowing for sample-efficient characterization of policy behavior across unseen conditions and a systematic identification of failure-prone regions. We conduct 2331 real-world evaluations across 3 tasks with 3 factor variations and find that our approach typically saves the evaluator at least 20-40% of trials compared to typical random testing.

Machine-learning datasets labelled "4D" universally denote three spatial dimensions plus time. We introduce HyperShadow, the first public benchmark in which the fourth, fifth, and sixth dimensions are spatial: the task is to decide whether a 3D point cloud is a native three-dimensional shape or the projection, the "shadow", of a rigid object living in R^N (N = 4-6). We show this task is fundamentally distinct from intrinsic-dimension estimation: a shadow is still at-most-3-dimensional data, and standard estimators (TwoNN, Levina-Bickel MLE) reach only 71-73% accuracy. Detection instead requires projection signatures, density folds, filled volumes with characteristic radial profiles, and topology changes, which a 190k-parameter point network recovers at 96.6% accuracy across four corruption tiers, generalizing at 79-91% to object families never seen in training. On a temporal track of rigidly rotating objects we introduce a zero-parameter rigidity witness: the residual of the optimal rigid 3D alignment (Kabsch) between consecutive frames, which must vanish for any rigid 3D motion but cannot vanish for the shadow of a rigid rotation in R^N. This single interpretable statistic separates the classes at AUROC 0.982. All data are generated reproducibly from seeds; the dataset, models, and code are released publicly. HyperShadow makes no claim about physical reality; it is a controlled instrument for studying which observable statistics can certify incompatibility with a purely three-dimensional explanation.

Ad-load design is a central supply-side decision in sponsored search: more sponsored slots can raise revenue, but may crowd out organic results and degrade user outcomes. We study this trade-off using a large-scale randomized field experiment on an Android app store, where over five million users are exposed to one through six sponsored slots. Increasing ad load raises revenue by up to 43%, but reduces total search conversions by up to 5% and daily engagement by up to 2.2%. These average effects mask substantial heterogeneity: additional slots generate large revenue gains for high-ad-conversion queries, but little or negative marginal revenue for low-conversion queries. The trade-off also shifts within query as advertiser composition changes, such as brand-advertiser presence. Motivated by these findings, we design and deploy a novel adaptive algorithm -- exploration-augmented Locally Adaptive Ad Load (e-LAAL). e-LAAL combines LAAL, a model-free query-level decision rule that updates ad-load recommendations using recent outcomes, with static exploration arms that maintain support and provide fixed-policy counterfactual benchmarks. We provide a finite-time dynamic-regret guarantee for the e-LAAL architecture. In a platform-level production deployment serving 22.3 million users and 77.6 million searches, e-LAAL improves the empirical revenue--conversion trade-off relative to deployed static benchmarks and outperforms uniform and historical query-dependent static benchmarks.

Spatially resolved omics studies increasingly combine transcriptomic and epigenomic assays, yet downstream analysis is often still performed using single-modality pipelines. We present LATTICE (Latent Alignment of Tissue-level and Transcriptomic Information for Cross-modal Embedding), a graph-based self-supervised framework that learns spot-level representations from harmonized multimodal features. LATTICE integrates five aligned modality blocks per Visium spot: Visium RNA, scMultiome RNA, scMultiome ATAC, spatial ATAC, and spatial CUT\&Tag. These modalities capture spatial transcriptomic measurements, single-cell inferred regulatory activity, and in situ chromatin and histone states within a unified lattice representation. LATTICE constructs a spatial neighborhood graph and trains a TransformerConv encoder using masked reconstruction, cross-modal alignment, and spatial smoothness objectives. On a private 11-sample melanoma cohort from an anonymized clinical collaborator comprising 54{,}912 total spots, LATTICE demonstrated stable optimization behavior, reproducible embeddings across analysis seeds, and complete multimodal integration across all samples. Adding scMultiome RNA to Visium RNA alone substantially improved concordance with Space Ranger clusters across 11 runs (adjusted Rand index [ARI] +0.157, normalized mutual information [NMI] +0.143, and spatial contiguity +0.174). Additional modalities further improved spatial contiguity and multimodal utility score (MUS), although they sometimes reduced agreement with RNA-derived reference labels, likely because the learned embeddings captured chromatin and regulatory structure beyond transcriptomic similarity alone. These results position LATTICE as a practical and empirically grounded framework for multimodal spatial omics integration, while also highlighting the need for stronger supervision and broader external benchmarking.

Many signal processing systems ultimately exist to {act}. Whenever the state variable that determines the action to be taken by a decision maker, or agent, is uncertain, the way that uncertainty is represented decides how well the agent performs and how much its performance can be trusted. This lecture note develops, from first principles and within a single decision-theoretic setting, the link between the {objective} and the knowledge of an agent and the form of uncertainty representation that is sufficient to act optimally. To start, assuming a known environment distribution, we show that a risk-neutral agent needs the posterior distribution over the state, whereas a risk-averse agent can rely without loss of optimality on a {prediction set} and a worst-case decision rule. We then turn to the case in which the environment is unknown, and identify three complementary approaches to address the resulting epistemic uncertainty: calibration of a fixed predictor, credal (ambiguity) sets with distributionally robust optimization, and Bayesian inference over model parameters. The common thread is that reliable decisions require an uncertainty representation matched to the decision objective and to the knowledge profile of the agent, together with a guarantee that certifies the utility the agent will actually obtain.

This paper describes DS@GT ARC's submission to the CLEF 2026 LongEval Task 4 on Retrieval-Augmented Generation (RAG). In this submission, we examine a divergence between traditional natural language evaluation metrics and citation integrity as applied to RAG QA systems. We evaluate a corrective pipeline using Corrective RAG (CRAG) and CiteFix against baseline and frontier model benchmark RAG QA scores. While frontier models maximized answer relevance and fluency scores, our RAGAs LLM-as-judge diagnostics indicate that frontier models would correctly identify relevant documents without using their context in answer generation. Conversely, by filtering chunks pre-generation and enforcing strict entailment of generated claims to the cited material post-generation, our corrective pipeline marginally improved citation faithfulness and answer grounding. We propose that evaluation of trustworthy RAG QA requires metrics that reward strict answer grounding.

Evaluations of language-model honesty read the model's verdicts as evidence about the model. We test the instrument instead. We built a text-adventure world where the game engine, not any model, knows whether the quest can be completed. A language model plays under a budget and must eventually declare its quest complete, unreachable, or not yet decidable; the engine scores every verdict. Decision rules were recorded before results were read, and run artifacts bind the revisions they executed; the strength of preregistration varies by series and is disclosed. With the player held fixed, instrument choices substantially changed measured behavior. On four byte-identical anchors, expanding a two-verdict grammar to three verdicts moved strong claims from 38/40 to 7/40, while the new incomplete verdict took 28/40 outcomes; across series 2, 93/158 valid games ended incomplete. One sentence disclosing the success criterion took matched-instance false verdicts from 18/59 to 0/58, through fewer decision points and cleaner decisions. Repeated runs of one fixed configuration produced non-stable verdict distributions on 3 of 4 instances: single runs report samples as dispositions. A formally preregistered narrative-register gradient was falsified; two post-hoc, hypothesis-generating patterns remain: register presence roughly doubled strong claims, and budget rendering moved verdicts more than register content (.383 meter vs .150 lantern). The narrator compressed abundant budgets toward scarcity landmarks, yet the registered mediation test returned a null. We propose a four-check integrity protocol for eval instruments.

Product catalogs are the backbone of e-commerce sites, yet a large number of structured attributes (SAs) -- such as material, color, and shape -- often have missing values. Typically, SA values are extracted from product information, including titles and descriptions. While LLM-based generator-evaluator frameworks have demonstrated effectiveness for SA prediction -- where an LLM generates SA values and another evaluates them -- they face challenges when the Generator and Evaluator produce conflicting outputs, as either component can make mistakes. We introduce \texttt{CatalogAgent}, a novel agentic system that continuously improves Generator and Evaluator models for e-commerce catalog enrichment. When disagreements arise from (1) internal conflicts between the LLM-based Generator and Evaluator, or (2) external feedback from sellers on LLM outputs, a Supervisor Agent intervenes to mediate these conflicts and make final decisions. The system also incorporates a Memory Base and a Memory Summarizer that stores Supervisor Agent activities from individual cases and aggregates patterns into learnings. These learnings are fed back to the worker Generator and Evaluator LLMs, enabling self-improvement without human intervention. Through context engineering -- injecting learnings and insights into worker LLMs' contexts -- the system successfully transfers the Supervisor's capabilities to the Generator and Evaluator, improving their performance by 15.24\% and 13.98\%, respectively. Our experiments demonstrate a new paradigm of Supervisor Agent-mediated self-learning systems for improving generative AI model accuracy.

Large language models achieve strong scores on medical benchmarks, yet these benchmarks evaluate each question in isolation, providing no measure of whether a system can distinguish clinically similar presentations requiring different interventions. We introduce MamaBench, the first counterfactual benchmark for maternal and paediatric AI: 434 expert-authored clinical narratives in 217 pairs across 371 pathologies, evaluated via the Bias Trap Rate (BTR), the conditional probability that a model fails the counterfactual given success on the base case. We propose Evidence-Anchored RAG (EA-RAG), a three-stage retrieval method that replaces aggregate similarity with an evidence coverage objective through clinical parameter extraction, coverage auditing, and contrastive sub-queries. Across eight configurations of four frontier LLMs, base accuracy overstates robust accuracy by 16-28 percentage points in every model. EA-RAG achieves 20.3% BTR and 65.0% robust accuracy on Claude Sonnet 4.6, a 5.5 percentage point BTR reduction without degrading base accuracy. The residual 20% BTR confirms that counterfactual robustness in clinical AI remains an open challenge. Keywords: counterfactual evaluation, clinical AI, maternal healthcare, retrieval-augmented generation, diagnostic robustness

Vision Language Models (VLMs) demonstrate strong perceptual abilities but remain limited in tasks requiring analytical reasoning across multiple visual states, such as multi-image comparison, change detection, and multi-step visual inference. These capabilities are critical for real-world multimodal applications where reasoning must be grounded in systematic differences between visual contexts. However, existing benchmarks rarely require both explicit visual comparison and analytical reasoning, leaving this capability underexplored. To address this gap, we introduce SD-MAR (Synthetic Data for Multi-image Analytical Reasoning), a framework for training and evaluating VLMs on multi-image analytical reasoning. SD-MAR constructs paired visual scenarios through controlled perturbations and generates reasoning tasks spanning semantic change attribution and quantitative comparison. We further train VLMs using GRPO-lite with Backward Discounted Allocation (BDA), a reinforcement learning approach that removes KL regularization to encourage stronger policy optimization while allocating greater credit to the later reasoning steps where analytical conclusions are formed. Experiments on Qwen2.5-VL-7B and InternVL3-8B show that GRPO-lite fine-tuning on SD-MAR improves in-domain accuracy by up to 36.95%, with Qwen2.5-VL-7B outperforming GPT-4.1 on the SD-MAR benchmark. Importantly, out-of-domain generalization is preserved or improved: performance remains within 1% on MME, MMMU-Pro, and MathVista, while improving by up to 4% on MMBench. LLM-as-judge evaluation further demonstrates consistent improvements in logical coherence and explanation quality across both models.

Efficient long-context inference is not only about reducing memory cost, but also about keeping useful contextual evidence accessible as generation proceeds. However, existing compression-oriented approaches, such as key-value (KV) cache compression and context compression, often either make an early decision about which contextual information to keep or rely on an external compressor. Such designs make it difficult to adapt the compressed context to the evidence needed by later reasoning steps. This paper introduces PReM (Preserve and Refresh Memory), a context-compression framework that maintains the long context as the model's internal layer-wise KV memory and learns what to preserve and when to refresh it. Specifically, PReM uses a dedicated memory layer to make memory-selection decisions, and a special memory token <m> to trigger refreshes during generation. To train this behavior, PReM introduces Phase-Separated Refresh Training, aligning memory selection with memory-conditioned generation while preserving continuity across refreshes. Experiments with 32K-token contexts show that PReM outperforms strong baselines under both 16x and 32x compression, while maintaining a favorable balance between answer quality and inference efficiency.

In this paper, we present a comprehensive framework for assessing the explainability of various XAI methods, such as LIME and SHAP, across multiple datasets and machine learning models, with the ultimate goal of creating a unified multidimensional explainability score. Our methodology focuses on three key aspects of explainability: fidelity, simplicity, and stability. We leverage benchmarking experiments to systematically evaluate these aspects and use the insights gained to construct an offline knowledge base. This knowledge base captures the explainability scores for each registered model and serves as a valuable resource for context-dependent evaluation of explainability. By analyzing the complementary characteristics and metadata of AI models, datasets, and XAI methods, the knowledge base will enable the estimation of explainability scores for previously unseen datasets and models. Properties like fidelity, simplicity, and stability may vary significantly based on the dataset, underlying model, and domain expertise of the end user. We demonstrate our framework by applying it to three open-source datasets, discussing the implications of the obtained results in relation to the characteristics of the datasets. Our work contributes to the growing field of XAI by providing a robust and versatile tool for evaluating and comparing the explainability of various XAI methods, ultimately supporting the development of more transparent and trustworthy AI systems.

Large language models are increasingly deployed as agents, but reliable agentic behavior requires more than next-token prediction. At inference time, it is preferred that an agent can decide whether to proceed with its current reasoning, defer to a stronger model, request additional information, invoke external tools, or abstain under the given setup. Existing approaches address these decisions through prompt-level routing, external orchestration, or task-specific fine-tuning, which primarily rely on input-side signals, and are often costly and difficult to maintain as model backbones evolve. We ask whether such control decisions can be inferred directly from a model's latent generation process. We introduce Multi-Head Latent Control, a lightweight layer that reads hidden-state trajectories from a frozen LLM or VLM to produce deployment-time control signals. A Capability Head predicts whether the current model can solve the instance or should defer to a stronger collaborator, while a Resolution Head predicts appropriate resolution decision Clarification, Tool Use, Abstention, or Direct Answering. Both heads are trained only on latent traces from the same frozen LLM backbone, enabling post hoc adaptation without modifying the model. Across language and vision-language settings, Multi-Head Latent Control consistently improves the quality-cost tradeoff of multi-model systems, enabling early handoff from partial generations and more accurate intervention decisions. In routed execution (small + large model), it reduces large-model usage by up to 90.7 percent on AndroidWorld and 27-53 percent on average across benchmarks, while retaining most of large-model performance. Additionally, the learned control signals improve tool-use decision quality, yielding up to +158 percent relative score gain and 65.5 percent fewer missed-required tool calls.

Context engineering has become central to building reliable AI agents, yet it remains largely unmeasured. Agents do not fail in isolation: their behavior is shaped by the instructions, tools, memory, retrieved knowledge, guardrails, and untrusted inputs accumulated in their context. When this context is weak, agents drift, hallucinate, misuse tools, ignore constraints, become vulnerable to injection, and waste tokens. This paper validates context-engineering quality as an independent leading indicator of agent reliability. We implement the measurement in ProofAgent-Harness, an open-source infrastructure for AI agent evaluation that uses multi-juror, consensus-based scoring. The harness assesses context across seven criteria: role clarity, guardrail coverage, instruction consistency, tool schema quality, grounding sufficiency, injection hardening, and token efficiency. Crucially, the context score is isolated from behavioral metrics and release decisions, enabling a non-circular validation. Through a controlled context-quality study across regulated agent domains, holding frontier LLM agents fixed and varying only their operating context, we show that context-quality criteria consistently predict their corresponding behavioral outcomes. Grounding sufficiency predicts hallucination resistance, guardrail coverage predicts manipulation resistance, instruction consistency predicts instruction following, and tool-schema quality predicts tool use. These findings establish context measurement as a validated preflight signal for agent reliability and position context engineering as an auditable layer of agent evaluation and governance.

Automated chest CT report generation remains challenging because clinically faithful reporting requires both whole-volume understanding and accurate description of localized anatomical findings. Here we developed and retrospectively evaluated MonteRET, a region-aware retrieval-enhanced framework for generating chest CT findings sections. MonteRET integrates global CT features with region-level anatomical representations, retrieves clinically relevant knowledge using predicted medical conditions and region-level vision-language alignment, and refines initial reports through a knowledge-guided report rewriting agent. We trained our model on a public cohort with 24,128 CT scans from RadGenome-ChestCT. We evaluated MonteRET on the public RadGenome-ChestCT test set of 1,564 CT scans and an external cohort of 82 CT scans from NewYork-Presbyterian/Weill Cornell Medical Center. MonteRET improved report quality, semantic similarity, and clinical efficacy compared with a matched baseline and several state-of-the-art methods. Gains were most pronounced for recall, suggesting fewer omitted findings. Human expert evaluation by radiology residents also favored MonteRET.

Multimodal Large Language Models (MLLMs) are increasingly deployed for nuanced content safety and moderation tasks, yet they remain vulnerable to adversarial attacks and out-of-distribution edge cases. Traditional active learning and manual annotation fail to scale against the complexity and volume of novel multimodal threats. In this paper, we propose an automated, agentic red-teaming framework that systematically synthesizes difficult examples using an iterative strategy that proposes novel hypotheses as well as mutating on past attempts. Leveraging a multi-agent architecture that consists of a high-reasoning Architect agent, an advanced image generator, and a multi-level verification committee of LLM raters, our system autonomously uncovers boundary-pushing violations and ambiguous policy edge cases without any human intervention. By employing these carefully synthesized adversarial examples as in-context demonstrations via test-time Retrieval, we substantially improve the target model's robustness, reducing the False Negative Rate (FNR) from 41.2% to 24.5% in a public image safety benchmark without relying on any human labeling.

Long-horizon robot planning requires more than predicting what actions will do next; it also requires memory of the embodied experience that makes future goals interpretable. People do not plan from the present scene alone: they draw on remembered places, object-state changes, prior procedures, and regularities revealed through repeated action. We formulate Embodied Action Memory (EAM) as the capability to form, maintain, and use such experience as a persistent memory state for later decisions. MEMORA realizes EAM with a formation-consolidation-retrieval lifecycle and four typed stores: Environment Memory, Entity Memory, Activity Memory, and Inferred Knowledge. Online editing maintains object identities and state histories as new observations arrive; offline consolidation abstracts repeated experience into reusable procedures and participant-specific regularities. MEMORA-Bench evaluates this lifecycle on 45 hours of EPIC-KITCHENS-100 extension video across 18 participants through memory-grounded planning, including previously unseen goals, and a complementary memory-assessment task. Across four open-weight language models, full MEMORA--combining editing, typed stores, and consolidation--achieves the strongest aggregate results among the evaluated memory conditions. It improves memory-assessment accuracy by up to 20.5 points over the strongest controlled baseline and improves out-of-distribution Robot-Grounded Plan score by up to 16.6% relative. A qualitative two-task robot deployment study further illustrates how memory-grounded language plans can interface with downstream control, while the overall results show that editable, consolidated memory can supply remembered context for robot planning. Project page: https://yuzihaowashu.github.io/MEMORA/

Automatic translation systems, from CAT tools to MT, overwhelmingly treat translation as a sentence-by-sentence act. This paper asks whether LLMs can be moved beyond that paradigm through whole-document, corpus-informed translation. We present PAT (Pragmatic Auto-Translator), a RAG-based system that pairs user-configured specifications with context from a comparable corpus of authentic longform texts in U.S. English and Latin American Spanish, passing retrieved paragraph-, section-, and document-level examples to an LLM for whole-document generation. The goal is draft translation for professional verification: target texts reformulated to fit their Spanish-language context, where discourse organization, rhetorical style, and pragmatic norms differ meaningfully from English. We evaluated six automatic translations of essays on generative AI across three projects using a customized MQM typology, assessed by two trained evaluators working from U.S. English into LATAM and Mexican Spanish. Results show that a limited prompt produced no meaningful reformulation, and specifications and corpus-informed translations at times showed substantial reformulation, though not always to effect. We find that LLMs can be moved toward reformulation and away from the sentence-by-sentence paradigm, though more work is needed to improve the effectiveness of those reformulations. In this paper, we discuss considerations related to automatic translation system design, corpus construction, and translation quality evaluation methodology and results.

Most reported gains from agent-optimization methods are one-shot: an agent is optimized against a fixed benchmark and the resulting improvement is reported as if it were a stable property of the method. This does not test the setting that matters for deployed agents, where optimization is applied recursively as new failures and new tasks appear over time. The central question this raises is whether optimizer-driven gains compound: after an agent has been optimized once, can it be optimized again on newly arrived tasks without eroding the gains the first round produced? We study this question with a two-phase continual-learning evaluation built from hard tasks in Terminal-Bench 2.0, comparing three approaches to agent-harness optimization (GEPA, Meta Harness, and RELAI's Verifiable Continual Learning, RELAI-VCL) under identical optimization budgets. All three methods improve over the baseline agent in the conventional, static, single-phase setting. However, once new tasks are introduced, the methods diverge sharply: GEPA's optimized agent transfers below the unoptimized baseline, Meta Harness transfers well but fails to improve further once given a second optimization budget, and RELAI-VCL is the only method that both transfers positively to unseen tasks and continues improving after those tasks are folded into the optimization objective, reaching the highest pass rate at every evaluated stage and the highest lifelong average pass rate overall (76.4% vs. 66.0% for GEPA, 64.6% for Meta Harness, and 58.7% for the baseline). Our key observation was that optimization gains compounded only when regression control was built into the optimization loop, providing an inductive bias against shortcut solutions that fail to generalize.

The rapid proliferation of Agentic Artificial Intelligence fundamentally disrupts traditional customer loyalty paradigms. As AI evolves from passive recommendation algorithms to autonomous, goal-directed agents capable of executing purchasing decisions, the conventional understanding of consumer-brand relationships requires a structural reevaluation. By synthesizing extant literature across human-machine teaming, consumer decision-making, and algorithmic trust dynamics, we demonstrate that traditional loyalty models fail to account for algorithmic bounded rationality and constructed autonomy. To address this, we introduce the Dynamic Verifiable Multi-Agent Human Agentic Loyalty Loop (DVM-HALL) model. We formalize brand choice via a softmax probability formulation where human emotional equity, agentic machine-experience utility, calibrated trust, delegated authority, and verifiable execution jointly determine selection. The model features recursive updating mechanisms to dynamically calibrate trust and delegation after each interaction. Crucially, the framework integrates a verifiable execution layer for Decentralized Finance (DeFi) and tokenized loyalty settings, incorporating execution risks -- such as gas costs, slippage, MEV exposure, and smart-contract vulnerabilities -- as core predictors of agentic brand preference. Furthermore, we introduce the Net Human-Agent Score (NHAS), an auditable, risk-weighted metric designed to measure human-agent alignment using human feedback, execution logs, benchmark comparisons, and verifiable receipts. Finally, we propose a comprehensive three-stage empirical validation plan spanning controlled shopping experiments, multi-agent market simulations, and DeFi testbeds. This framework provides the foundational theory required for brands to navigate the impending transition toward machine customers.

Large language models are increasingly deployed in agricultural decision-support settings, yet high-stakes crop protection in smallholder agriculture requires more than output-quality benchmarks. Over a two-year design and evaluation programme, we formalise policy-constrained large language model assessment as an adaptive compute allocation problem that jointly captures safety compliance, helpfulness, operational latency, and expert supervision workload. We introduce P-EVAL (Policy-grounded Expert-calibrated VALidation protocol), a unified evaluation framework for policy-grounded decision support, evaluating the architecture on a simulated field query database consisting of 1,240 cases. The protocol is instantiated on the Pezego advisory architecture (Pezego-HITL) and evaluated in Ghana. Following offline judge calibration against gold-standard human expert decisions ($κ= 0.77$), we evaluate the architectural performance under simulated query workloads. Under P-EVAL, our memory-routed architecture improves the Policy Alignment Rate (PAR) to 0.94 and the Agronomic Utility Rate (AUR) to 0.95, while reducing P95 latency by 55% (from 28.6s to 12.9s) through a 59.6% cache reuse ratio. We also demonstrate generalisability using the open-source \texttt{Qwen3.5-9B-DeepSeek-V4-Flash} model, achieving a PAR of 0.86 and a 54.5% latency reduction (to 10.2s). To evaluate practical utility and socio-technical integration, we administer detailed questionnaires to Ghanaian Extension Services Officers ($N=30$) and smallholder farmers ($N=36$). Taken together, this work demonstrates how policy-grounded structured retrieval-augmented generation with validated-memory routing makes safety-utility-latency trade-offs explicit, offering a scalable template for trustworthy AI-driven extension in smallholder farming systems.

While search agents demonstrate impressive capabilities in multi-step question answering, their robustness to poor-quality evidence remains under-explored. This phenomenon occurs rarely in realistic benchmarks but can lead to dramatic failure in real life applications. Therefore in this study we propose DeepStress, a stress testing framework that controls the frequency of challenging evidence by replacing the retrieval module of search agents with a controlled synthetic environment. We use this framework to control three dimensions that can affect document reliability: trustworthiness, relevance, and factuality. Testing several search agents on HotpotQA and BrowseCompPlus, we demonstrate that agents exhibit substantial differences in their ability to handle unreliable information and propose new metrics that better document systems outcomes as well as the interactions between conflicting parametric and retrieved knowledge.

Multimodal agents that think with images iteratively manipulate visual evidence and invoke tools across many steps. Existing reinforcement learning methods reduce trajectories to scalar rewards, forcing the policy to discover reusable tool-use patterns from scratch on every new task; memory-based alternatives retain past experience, yet they rely on test-time retrieval, without updating the policy to absorb reusable patterns from that experience. Our key insight is that multimodal reasoning trajectories should be distilled into reusable skills that co-evolve with the policy during training, rather than being consumed as rewards or retrieved from a static store. To this end, we propose SPyCE (Skill-Policy Co-evolution), a framework that distills trajectories into a hierarchical skill library and updates it throughout reinforcement learning. Execution skills capture local visual operations, while workflow skills encode high-level priors that orchestrate tool use. During training, the policy model conditions on retrieved skills to guide its rollouts, while the skill library evolves using valuable rollouts generated by the policy. This creates a closed loop in which improved policies yield better skills, and the evolving skill library, in turn, provides stronger priors for policy rollouts. Experiments across eight benchmarks demonstrate that SPyCE consistently outperforms both RL-based and memory-based baselines. Further analysis reveals that both the hierarchical skill design and the co-evolution mechanism are critical to our design. These results suggest joint skill-policy optimization as a promising paradigm for building capable multimodal agents.

Recent advances in Large Language Models have fueled autonomous AI agents capable of tackling complex scientific tasks, yet existing automated research systems remain predominantly focused on empirically driven domains with quantitative benchmarks, leaving theory-driven discovery, particularly in mathematically grounded disciplines requiring rigorous proofs and synthesis of domain knowledge, largely underexplored. Key challenges include the difficulty of verifying theoretical reasoning at scale, insufficient reasoning ability for autonomous frontier exploration, and a scarcity of procedural heuristics in the literature. We introduce ReasFlow, an end-to-end autonomous agent system for reasoning-centric scientific discovery that operationalizes a collaborative paradigm where the human expert acts as Principal Investigator while the agent executes rigorous derivations as a capable graduate student. ReasFlow incorporates (i) a robust internal verification loop that audits logical coherence and corrects fundamental errors prior to human inspection, and (ii) an automated knowledge retrieval and self-improvement mechanism that proactively surfaces both declarative facts and overlooked procedural heuristics, substantially reducing expert intervention. The system unifies literature synthesis, algorithm design, theorem proving, experimentation, and manuscript preparation in a single system. Deployed to autonomously generate five complete research papers with rigorous theoretical and empirical content from minimal prompts, ReasFlow consistently achieves the highest evaluation scores among state-of-the-art open-access baselines under a curated LLM-based review rubric. ReasFlow is publicly accessible via the ReasLab platform, providing a collaborative workspace for AI-assisted theoretical research. Github repo: https://github.com/ReasLab/ReasFlow.git.

Large language models have made strong reasoning gains through supervised fine-tuning, reinforcement learning, and on-policy distillation, yet these post-training methods are usually evaluated only by final-answer accuracy. We study how they reshape confidence during reasoning. We introduce a three-stage calibration framework that evaluates confidence before, during, and after chain-of-thought generation, corresponding to difficulty estimation, early termination, and answer aggregation. Through a controlled comparison on mathematical reasoning benchmarks, we find that OPD provides the most useful pre-reasoning confidence, SFT gives the strongest online signal for early stopping, and RL produces the most reliable trace-level signal for aggregation. We further show that confidence reliability is position-dependent: RL confidence becomes informative after a path-commitment phase, while OPD confidence is useful early but can become inversely calibrated later. Based on this observation, we propose PosConf, a position-aware confidence strategy that uses confidence only from reliable relative-position intervals. PosConf improves RL answer aggregation by 6.1 points over majority voting and consistently improves OPD early stopping under tight token budgets, with gains up to 4.3 points by avoiding its later inverse-calibration region, showing that \emph{confidence in reasoning models should be used both stage-wise and position-awarely}. Our code is available at https://github.com/EIT-NLP/Post-Training-Calibration.

Studies of bias in LLM-as-judge systems typically build synthetic corpora by prompting an LLM to generate a hallucinated answer to pair with a factual one, then presenting both to a judge. We report a case in which this generation step silently failed, and use it to argue that the failure mode is structural rather than incidental. In a multilingual (Turkish/English) faithfulness-judgment corpus, a decoding-budget parameter shared between judging and generation calls truncated one producer's hallucinated answers to a few words. The resulting items produced a large, statistically robust effect: a 32-point cross-lingual collapse in one judge's selection accuracy, replicated from N=50 to N=500, explained by a three-layer mechanistic account, and confirmed by a controlled producer-swap experiment, none of which was real. The effect vanished to ceiling once the shared parameter was corrected, and only manual reading of the raw generations, not any aggregate statistical check, exposed the fault. A second measured bias (Markdown-formatting preference) was not fabricated but distorted by the same fault, its magnitude and in one case its sign shifting with stimulus length, a mode aggregate metrics cannot distinguish from the first. We frame the underlying vulnerability using the test oracle problem: corpora whose negative examples are LLM-generated carry no mechanical way to verify item integrity, while corpora built by deterministic perturbation of a gold answer carry an item-level oracle for free. A positive control supports this claim directly: an analogous fault injected into a minimal perturbation-based corpus is caught with 100% accuracy by a zero-cost, zero-human gold-to-negative string comparison. We close with a validation protocol, derived from our own case, for analysts working in the oracle-less regime that we argue describes most contemporary multilingual LLM-as-judge corpora.

An LLM agent's real-task performance is shaped as much by the harness around its model as by the frozen model itself: its prompts, injected knowledge, runtime control, and configuration. In deployment the harness is often the only lever available, so improving it automatically is the natural way to raise performance without touching the weights. The hard part is not generating changes but knowing which one truly helped. Self-generated feedback is noisy, and an apparent gain can be a measurement artifact or an edit that merely overfits the tasks it was tuned on. We present a self-evolving agent-harness framework that separates proposing changes from crediting them: a language model diagnoses failures and proposes patches, while all sampling, measurement, and significance testing are owned by deterministic code, so every credited improvement is trustworthy by construction. Patches populate a gated, categorical quality-diversity archive (GSME) keyed on the (WHERE x WHY) pathology an edit addresses rather than the tasks it fixes, an anti-overfitting inductive bias; generalization is measured on a sealed test scored only after evolution. Across seven domains with a frozen open-weight model, the harness is train-selected and scored once on a sealed test; its credited gains there are +9 to +15.5pp and retain 86-147% of the training gain, evidence they generalize rather than overfit. The winning patch tracks the model's dominant pathology, not its size or family: changing the model can change the pathology and the patch, while the same pathology-to-patch match recurs across two model families. What transfers is the diagnose-and-credit loop, not any specific harness.

Sampling multiple solutions and returning the majority answer is among the most reliable ways to improve the reasoning accuracy of large language models without labels, and a growing family of methods converts this consensus signal into training supervision. However, existing approaches use consensus only in restricted forms: as a filter that selects solutions for fine-tuning, as a preference between answers, or as a scalar reward for reinforcement learning, discarding most of the information that the agreeing solutions contain. We present CANON (Consensus-ANchored self-distillatiON), a label-free training method that turns consensus into dense, token-level supervision. For each unlabeled prompt, CANON samples multiple solutions, extracts the majority answer, and conditions a frozen snapshot of the model on a solution that reaches it; this consensus-anchored teacher then supervises the model on its own rollouts at every token. Experiments on mathematical and scientific reasoning benchmarks show that CANON improves pass@1 by up to 12 points, outperforming label-free reinforcement learning by 6 points at a seventh of its compute and approaching a teacher conditioned on gold solutions; trained on pooled unlabeled data, it transfers to held-out benchmarks, matching training methods that use gold labels. Analysis suggests that the improvements are not pure distribution sharpening: after training, the model solves problems it previously never solved in 32 attempts, and its majority vote itself becomes more accurate.

Reinforcement learning has emerged as the dominant paradigm for training large language model (LLM) agents that interact with executable sandboxes. State-of-the-art algorithms such as PPO, RLOO, and GRPO inherit their rollout topology from RLHF: for each prompt, N independent trajectories are sampled from the initial state, and an advantage is computed by subtracting a group baseline. This design ignores a defining property of agent sandboxes. They are deterministic, snapshottable, and resumable from any intermediate state. We argue that this property enables a fundamentally different rollout topology: rather than N independent trees of depth T, one can construct a single tree of N leaves whose siblings share prefixes, and therefore share variance. We instantiate this idea as Branching Policy Optimization (BPO), a sandbox-native RL algorithm that (i) adaptively snapshots the sandbox at high-entropy decision points along a backbone trajectory, (ii) forks K alternative actions per branch point and rolls out each to termination, and (iii) computes per-step advantages from sibling returns rather than from independent prompts. We prove this estimator is unbiased and has strictly lower variance than the trajectory-level baseline, with the reduction equal to the prefix-explained portion of return variance. On WebShop, ALFWorld, and SWE-bench Verified with Qwen2.5-7B and Llama-3.1-8B backbones, BPO improves success by 3.6--6.1 absolute points over GRPO and RLOO at matched compute, halves gradient-norm variance, and matches the best baseline using 38% fewer policy updates.

Algorithmic collusion among pricing algorithms has raised concerns about sustained supra-competitive prices and their implications for social welfare. Existing work has largely focused on the probability that reinforcement-learning algorithms converge to cooperative strategies, typically under the assumption that exploration vanishes over time. Motivated by the observation that algorithms deployed in practice are likely to continue exploring in order to remain adaptive to changing environments, we study learning dynamics under constant exploration. In this setting, the relevant question is no longer whether an algorithm converges to a particular strategy profile, but rather what fraction of time the algorithms spend playing cooperative strategies. Even in the benchmark case of the repeated Prisoner's Dilemma with one-period memory, this yields high-dimensional stochastic learning dynamics, for which a complete analytic treatment is intractable. We show that cooperative strategies can be dominant in this time-averaged sense and derive a boundary predicting when such dominance arises, based on the expected dynamics of the Q-learning process. Extensive simulations show that this boundary is a strong predictor for non-defection-dominated behaviour under epsilon-greedy Q-learning.

Systematic comparisons between current situations and structurally similar past events in the historical, i.e., historical analogies, is among the most powerful tools for foresight analysis. In this work, we present a new task called Analogical Deep Research (ADR) to Large Language Model (LLM) agents and construct the first ADR benchmark ADR-bench to study whether LLM agents are able to find and leverage historical analogies when doing foresight analysis. Our investigation reveals a key obstacle: LLM agents are poor at finding analogies because they match on surface features rather than underlying mechanisms. We argue that ADR is inherently a causal question as it requires understanding why the event occurred. Based on our theoretical analysis, we propose two principles required for ADR, including the mechanism alignment and cross-analogy confirmation. Built upon our theoretical results, we propose a new agentic framework called Causal Analogical Researcher (CANA) that guides LLMs to find and integrate historical analogies. CANA incorporates a simple yet effective structural decomposition representation, and integrates structural feedback for reflective improvements of historical analogy identification and integration. We show that CANA brings up to 10% improvements in historical analogy generation, and surpasses the state-of-the-art deep research agents in the ADR-bench. Case studies with the ongoing events confirm the effectiveness of CANA in leveraging historical analogies.

Large Language Model (LLM) agents increasingly rely on external memory systems to accumulate experience across tasks. Yet nearly all existing approaches, from graph-structured memories to reflective insight stores, access memory through fixed, hand-designed heuristics. We argue that this static view of memory is a core bottleneck for agentic learning because optimal memory behavior is fundamentally context-dependent. The early stages of the tasks, benefit from minimal retrieval because memory is sparse; recurring goal types benefit from plan reuse rather than generic nearest-neighbor lookup; stuck agents benefit from re-retrieval with alternative queries; and across long task streams, the memory store itself must be consolidated and pruned to remain useful. We present Memory as a Controlled Process (MemCon), a framework that models memory operations as a Markov Decision Process and learns an online policy that adaptively decides when, what, and how much to retrieve, when to inject a distilled plan, and when to consolidate or forget. MemCon is backend-agnostic: it wraps any existing memory implementation, learns from task-by-task binary feedback with no pretraining and no additional LLM calls, and uses a lightweight tabular contextual bandit with UCB exploration that converges within tens of tasks. Across 6 benchmarks, 3 agent frameworks, and 3 LLM backbones, MemCon consistently outperforms multiple memory baselines by up to 15.2 points in task success while reducing token consumption by 5--20%.

We investigate which language model evasion attacks survive state-of-the-art adversarial fine-tuning, developing strategies that sweep the top 5 positions on the ELOQUENT 2026 Voight-Kampff leaderboard. While adversarial fine-tuning trivially closes the 2025 winning evasion recipes, we uncover a fundamental asymmetry in detector vulnerability: pushing generated text out of the detector's training distribution reliably defeats adversarial detection, whereas pulling it into the distribution (e.g., mimicking human training data) fails completely. Exploiting this, we introduce two novel out-of-distribution attack families - cross-decade register attacks and modernist stream-of-consciousness form. Both strategies easily bypass adversarial closure, achieving up to approximately 50x higher fool rates than previous methods while preserving naturalness. Furthermore, experiments show that the obvious deployer countermeasure (augmenting training data with period prose) fails to close the vulnerability. Our findings show that the tested detector families, including adversarially fine-tuned ones, exhibit persistent vulnerabilities under structural out-of-distribution shifts, a mechanism that directly powers our leading competition performance.

We describe our BioASQ Task 14B 2026 system. The work centers on two design decisions: how aggressively to re-retrieve when first-stage retrieval is weak, and how to combine multiple language-model answers. Retrieval unions two parallel pipelines - a hybrid first stage (dense BGE + BM25 + RRF, reaching R@200 = 99.3% on the BioASQ-13b historical archive) and an agent-driven pipeline that decomposes the question over PubMed, Europe PMC, and iCite - with a BGE cross-encoder quality gate flagging weakly-supported questions for selective re-retrieval. On Task 12B 2024 validation, a cost-pragmatic re-retrieval policy beats a skill-strict baseline significantly on list F1 and list precision, at 12% lower re-retrieval cost. Holding prompt and model fixed across val and test 13B (different question sets), list F1 rises by +0.132 absolute on the BioASQ-released gold-input pool, consistent with substantial retrieval-side headroom. For Phase B answering we decompose multi-model ensemble lift into a selection component bounded by the per-question oracle and a fusion component that aggregators can exceed. The decomposition predicts before any experiment that LLM-as-judge wins on selection-dominated metrics (yes/no, multi-reference ROUGE) but is structurally insufficient on the recall component of fusion-friendly metrics (factoid rank-1, list recall). On Task 13B 2025 our synonym-union resolver wins list recall on every head, while GPT-5.5 solo retains the list-F1 lead because the resolver's wider item set costs precision. On the Task 14B 2026 preliminary leaderboard our team places first on the combined-exact aggregate on three of the eight (phase x batch) leaderboards, wins four individual question-type cells, and takes #1 on Phase B b3 ideal.

Large audio-language models (LALMs) are increasingly used as automatic judges for speech evaluation. However, high agreement with human ratings does not guarantee that their verdicts are grounded in the audio. A judge may instead rely on specialist labels or reference data supplied by the evaluation protocol itself, taking a shortcut in place of listening to the audio. In this paper, we audit such protocol-level ``shortcuts'' in LALM judges across three common deployment protocols: feature-blueprint judging, where the audio is replaced by a structured text description of acoustic features, reference-conditioned judging, and pairwise A/B comparison. Across six judges and four attributes, we find that several LALMs rely on protocol-level shortcuts. For example, in feature-blueprint judging, incorrect specialist labels reduce five judges' emotion accuracy to 0.10 or below, and in concatenated A/B comparisons, Qwen3-Omni-Thinking often picks the same slot regardless of order swaps. These results indicate that aggregate agreement can overstate the validity of LALM judges unless the model and the evaluation protocol are assessed jointly, and that each model-protocol pair should be evaluated with a matched shortcut probe.

We present MyAG, a graph-based framework for designing and analyzing composable LLM agent systems. Our framework separates agent system construction into three graph abstractions: a component graph for agents, environments, and modules; a workflow graph for execution control; and a search graph for runtime execution. This separation allows users to flexibly reuse the same components with different strategies. We further support hierarchical composition through recursive system nodes and provide monitoring and visualization tools for inspecting agent execution. Experiments on representative agent applications show that our framework supports flexible agent system design and helps analyze performance-efficiency tradeoffs. Our framework is publicly available and fully open-source.

LLM-based agents have rapidly improved at operating individual digital environments such as mobile applications, desktop systems, and smart homes. However, real-world user goals often span multiple devices: information may come from a phone, be processed on a desktop, and the result may need to appear on another device. Most existing benchmarks center on a single dominant execution environment, making it difficult to evaluate whether agents can acquire and integrate information across heterogeneous devices and complete end-to-end tasks with cross-device dependencies. We introduce DevicesWorld, a large-scale executable benchmark for cross-device collaborative operation. DevicesWorld contains 6,140 tasks and integrates three classes of device environments -- mobile, desktop, and IoT -- into a unified cross-device interaction and evaluation framework. Each task defines a natural-language user goal, participating devices and initial states, executable actions, rule-based verifiers, and a cleanup procedure. A multi-stage construction and quality-control pipeline keeps tasks close to realistic user needs while allowing final outcomes to be automatically verified from device states and generated files. We evaluate five frontier LLM-agent systems on a fixed evaluation set. All methods achieve low success rates, with the best reaching only 12.5%. Among failed runs, about 28.7% satisfy at least one scoring condition yet still fail the full task. Trajectories show that agents become stuck acquiring information or manipulating interfaces, confuse source and output devices, or terminate before all conditions are jointly satisfied. DevicesWorld turns cross-device collaborative operation into an executable, reproducible, and diagnostically useful evaluation problem for research on reliable cross-device agents.

We present a benchmark and reference system for live captioning of Sikh Kirtan - the continuous, sung recitation of verses from the Sri Guru Granth Sahib Ji (SGGS). Unlike open-vocabulary lyrics transcription, Kirtan captioning is a closed-vocabulary problem: every displayed line must be an exact, word-for-word line from the canonical scripture, because displaying misspelled Gurmukhi is considered religiously inappropriate. We formalize the task as predicting, at every time t, a pair (shabad_id, line_idx) or null, and organize the problem space into a 2x2 matrix along two orthogonal axes: live vs. offline (causal vs. full-audio access) and blind vs. oracle (shabad identity discovered vs. given). We release v1 of the benchmark - 4 hand-annotated Kirtan recordings x 3 cold-start offsets = 12 evaluation cases, ~57 minutes of scored audio - together with a scorer that computes frame accuracy at 1s resolution over a scored region, with a 1s collar and gap-tolerant scoring at segment boundaries. We describe a reference system (fine-tuned 120M IndicConformer -> fuzzy matcher -> state machine; INT8 ONNX; RTF ~0.05 on one Apple Silicon core) that achieves 57.9% overall frame accuracy across all 12 cases (10/12 correct shabad locks) on the hardest variant (live x blind). We compare against three trivial baselines (empty, shifted-5s, perfect) and discuss why standard ASR metrics (WER/CER) measure transcription accuracy rather than the display accuracy this task requires. The benchmark, reference system, and a live deployment are released under permissive licenses to facilitate further improvements.

Discrete denoising diffusion models (DDMs) have recently emerged as a compelling alternative to autoregressive (AR) modeling for discrete data, offering parallel generation and iterative global refinement capabilities. Unlike continuous diffusion, where the state space is fixed, DDMs are fundamentally shaped by how the discrete state space is constructed: the tokenization scheme, the vocabulary topology, and domain-specific structural alphabets. This work introduces a unified conceptual framework that views discrete diffusion models through the construction of the underlying discrete state space. Within this framework, existing formulations, including transition-matrix, masking/absorbing-state, and score/ratio-based approaches, emerge as different instantiations of a common design space. The framework further exposes common design trade-offs across training objectives, inference algorithms, scaling behavior, systems optimization, and evaluation protocols, suggesting several promising directions for future research.

Translating complex biomedical data into patient-friendly narratives is central to modern biomedical informatics. This study presents a comparative analysis of training small language models (SLMs) in specialized biomedical datato-text generation tasks. We explore widely adopted post-training methods including supervised fine-tuning (SFT), direct preference optimization (DPO), odds ratio preference optimization (ORPO), and group relative policy optimization (GRPO) with Qwen-based SLMs on a medicine package leaflets dataset. To assess cross-dataset generalizability, we also curated drug label data from openFDA. We evaluate models using both standard lexical overlap metrics like ROUGE as well as semantic similarity measures. Across our experiments, the results show that (1) the aligned SLMs outperform proprietary models like GPT-5; (2) ORPO outperforms the SFTbaselines; (3) GRPO yields the most robust cross-dataset performance among the alignment methods tested as well as GPT-5.

Recent text-to-audio models generate high-quality audio, but often fail to follow instructions involving multiple sound events and temporal order. This gap arises because existing evaluation and training signals mainly emphasize global similarity or perceptual quality, with limited supervision on instruction-level correctness. We propose an instruction-level framework that uses audio-aware large language models (ALLMs) as fine-grained judges to verify target event presence and temporal relations in generated audio. After validating ALLM judgments on benchmarks and through human verification, we use their feedback to construct preference pairs for direct preference optimization. We further introduce S3Bench, a narrative benchmark for evaluating multi-event temporal instruction following. Experiments show that our method improves event completeness, temporal ordering, and joint instruction-following accuracy across existing benchmarks and S3Bench, while maintaining audio quality.

On-policy distillation (OPD) has become a key paradigm in LLM post-training, yet its training dynamics remain poorly understood. We present a systematic study examining the role, pathologies, and regulations of OPD. We first clarify the role of OPD as an exploration catalyst: it steers the student toward correct reasoning paths via dense token-level guidance, without expanding capability ceiling. We confirm this by showing that prompt diversity matters more than per-problem sampling numbers, and critically, that the effectiveness of OPD hinges entirely on the quality of its guiding signal. This dependency exposes two pathologies that derail exploration. The Student-Teacher Mismatch occurs when a large teacher-student distributional gap causes the guiding signal to misalign with task correctness, steering exploration in counterproductive directions. Length Exploitation arises when the aggregated token-level objective creates length-dependent shortcuts, allowing the student to game the reward landscape through response truncation or redundant padding, exploring degenerate length modes rather than reasoning strategies. To tame these pathologies, we investigate lightweight signal regulations: advantage clipping and log-scale compression, ensuring exploration is guided by faithful signals. Experiments across seven benchmarks demonstrate that these regulations alleviate length exploitation and enable effective distillation, stably surpassing OPD variants and RLVR baselines, thereby confirming that well-regulated signal quality, rather than mere teacher scale, governs successful exploration in OPD.

What happens to an LLM agent's tool choice when the reliable tool silently changes within an ongoing session? We borrow set-shifting from cognitive psychology to study how well agents adapt to hidden reliability shifts. Our benchmark mounts tool-skill libraries with redundancies, where many tools solve the same task but differ in hidden reliability. In our evaluation framework, a branched schedule shifts the reliable tool group at hidden boundaries and pairs every shift with a no-shift control. We find that agents, by default, settle on a small recurring routine within a few turns of each boundary, with call shares concentrating on a few discrete values after each reliability shift. We score the set-shifting accuracy for each agent trajectory: the joint probability of routing to the target tool group in every post-shift window. We test open-weight LLMs in an open-source agentic harness and find qualitatively distinct failure modes across the same set of routines. We also find that set framing, how the toolset presents the alternatives as competing or complementary, shifts the routing dynamics.

Generative Flow Networks (GFlowNets) offer a promising alternative to reward-maximizing reinforcement learning (RL) for large reasoning models, encouraging diverse reasoning paths by matching reward distributions rather than collapsing to dominant modes. Recent work shows promise on math and code, but scaling GFlowNet-style RL to modern post-training pipelines remains difficult: as model size, rollout horizon, reward noise, and distributed-systems complexity grow together, a learned prompt-conditional partition function becomes a source of gradient instability and engineering overhead rather than a useful normalizer. Through systematic analysis, we find that the learned partition function, previously treated as essential, can be replaced by an in-batch Monte Carlo estimate computed from the rollout group already required for training. We propose GFlowRL, a streamlined GFlowNet-style RL algorithm that removes the auxiliary partition network entirely while preserving the reward-distribution-matching objective, completed by two stabilizers: importance-sampling correction for rollout/trainer drift and asymmetric flow-gap clipping for outlier residuals. GFlowRL exceeds all counterparts on math, code, and adversarial red-teaming benchmarks, reaching a Codeforces rating of 2048 at the 14B scale (within 25 Elo of o3-mini) and attaining the highest average ASR@1 on AdvBench and HarmBench, outperforming the previous SOTA multi-turn attacker in a regime where FlowRL, a prior GFlowNet-style method, diverges. The same recipe transfers to all evaluated MoE configurations up to 235B parameters, where FlowRL again fails to converge. To our knowledge, GFlowRL is the first GFlowNet-style RL algorithm to scale stably across both dense and sparse architectures. Code will be at: https://github.com/microsoft/gflowrl

Reinforcement Learning (RL) post-training is increasingly used to adapt foundation models for reasoning, planning, and feedback-driven robot-learning pipelines, but constrained post-training resources are often summarized by a single total FLOP budget. We study the fixed-budget decision problem behind this practice: under the same post-training budget, should one use a larger policy, train a smaller policy longer, generate more rollout search, or spend compute on stronger reward feedback? We introduce a FLOP-accounting framework for GRPO post-training that decomposes compute into rollout/search, policy-update/learning, and reward- or feedback-model evaluation. Across LoRA-adapted Qwen2.5 policies, we find conditional allocation frontiers: the best observed allocation changes with model size, compute budget, reward system, and evaluation target. Same-FLOP model-size comparisons show that model choice and training allocation are coupled because larger policies consume more per-token compute and therefore buy fewer updates or rollouts under the same budget. Reward systems also change the accounting: rule-based rewards spend nearly all non-update compute on policy rollouts, while PRM-style feedback allocates a visible part of the budget to reward-model inference. We present RACE as a diagnostic pilot-grid protocol, not a guarantee of held-out improvement, for identifying allocation regimes before expensive validation runs; our results suggest that RL post-training papers should report total FLOPs together with how compute is divided among model size, search, learning, and feedback.

Multi-agent systems (MAS) coordinate multiple LLM-powered agents through structured workflows, gaining reasoning power but incurring high inference latency from multi-step execution and repeated model invocations. Existing orchestration methods primarily optimize task performance and inference cost, leaving latency largely unaddressed. In MAS, end-to-end latency is governed by the critical execution path, so reducing total cost alone does not reliably reduce latency. Moreover, optimizing latency while preserving accuracy remains non-trivial: naive latency optimization can misassign operator-level credit and degrade task accuracy. To address this gap, we propose Latency-Aware Multi-agent System (LAMaS), a latency-aware orchestration framework for learning-based multi-agent systems. LAMaS addresses this challenge at two levels: at training time, it learns latency-aware execution graphs through constrained optimization with critical-path-aware credit assignment; at inference time, since a graph committed at training time cannot exploit runtime evidence, it complements graph construction with a lightweight controller that adaptively eliminates redundant future agent interactions as execution unfolds. Experiments on four benchmarks show that LAMaS achieves the best latency among evaluated learning-based MAS baselines, reducing end-to-end latency by over 50\% while maintaining competitive or better accuracy. LAMaS is also modular and transfers to other MAS with minimal changes, consistently yielding latency reductions.

LLM-as-a-judge is widely used to provide feedback and selection signals in closedloop regeneration, but this use remains insufficiently validated. We study it in table recognition, where deterministic TEDS evaluation provides a controlled testbed, using FinTabNet and OmniDocBench. Three findings emerge. First, judge signals were weak on both datasets: scores frequently tied, rankings were not reproducible, and the only selection policy that beat random on both datasets depended on an earliest-iteration tie rule, so its advantage cannot be attributed to the judge scores alone. Iteration produced better candidates, but the judge failed to recover them. Second, severe losses occurred even without specific judge feedback. A structurepreserving instruction significantly reduced the severe-loss rate on FinTabNet and was directionally consistent on OmniDocBench. The contrasts support target-preservation failure under unconstrained regeneration as a proximate mechanism of the observed severe losses. Third, the structure-preservation constraint reduced the severe-loss tail but produced no improvement. In an exploratory 2x2 analysis, the same protection was not stably observed when judge feedback was retained. These results do not dispute the value of LLMs as evaluators. Instead, they show that evaluation ability does not imply optimization utility. Iterative refinement requires, at minimum, a verification signal that deterministically detects structural change, rather than judge scores alone.

This position paper explores how Agentic AI and Model Context Protocol (MCP) can support power-grid studies in a Transmission System Operator (TSO) context. We focus on integrating Large Language Models with numerical simulation tools, structured workflows, and human supervision. We identify key industrial requirements for agent assisted grid studies and introduce pypowsybl-mcp, an MCP-based interface exposing selected capabilities of our simulation tool, pypowsybl to AI agents. This first step provides a testbed to study how agents can setup simulations, execute analyses, retrieve results, and interact with power-system simulators through standardized tool calls. We also discuss principles for human-in-the-loop, multi-agent workflows and outline an evaluation strategy combining technical metrics and practitioner feedback. The paper positions MCP-based tool integration as a step toward more interactive, auditable, and scalable grid-study environments.

The stable marriage problem appears in many privacy-sensitive domains, for example in the National Resident Matching Program in the US. In such applications, preserving the privacy of users' preference lists is essential to prevent strategic manipulation, discourage misreporting, and comply with data protection regulations. In this work, we investigate privacy attacks on stable marriage algorithms. Assuming that the attacker (e.g., the hospitals) can repeatedly interact with the stable marriage algorithm, we demonstrate how such interactions can reveal private preferences of the non-malicious side (e.g., the residents). We show that the widely applied Gale-Shapley Matching Algorithm, where the proposers' side is malicious, is vulnerable to privacy attacks and all honest agents' preferences can be revealed. We further investigate which preference distributions of the honest, non-malicious side are susceptible to privacy attacks and show that the Gale-Shapley Matching Algorithm where the honest side proposes can preserve privacy in non-susceptible preference distributions. We extend our results to the decentralized setting and show that the attacker's side can infer all preference orderings. In an experimental evaluation, we test privacy attacks on synthetic and real-world data and show that real-world data is indeed susceptible to privacy attacks. This work underlines a need for new privacy-preserving stable marriage algorithms.

As LLM technology advances, the space of model families, compute hardware, quantization schemes, parallelization strategies, and specialized optimization kernels continues to expand, sharply increasing the code complexity and maintenance cost of general-purpose inference frameworks. Conventional software engineering uses multiple layers of abstraction to support diverse application scenarios, but these abstractions also increase system complexity and may introduce additional performance overhead. This paper presents metainfer, an 'LLM-as-Compiler' approach in which users specify only the runtime constraints of an inference program. An LLM-driven multi-agent collaboration system, coupled with a contract knowledge base, then automatically generates a compact customized inference framework that satisfies these constraints. We evaluate metainfer from three perspectives: the effect of source-code reference, the runtime behavior and performance profile of engines generated under the zero-reference constraint on CKB-covered targets, and knowledge-base evolution for new model and platform scenarios. The results show that metainfer organizes generation constraints, validation feedback, and knowledge consolidation into a continuous closed loop, enabling runnable customized inference solutions to be generated from explicit knowledge. The code is publicly available at https://github.com/MetaInfer/MetaInfer.

Self-evolving agent systems improve by creating, revising, and retiring their own skills, but every such loop rests on a hidden assumption: a reliable evaluation metric already exists. In many real applications it does not. We make three claims. First, metrics can be \emph{evolved}: our metric loop searches compositions of small drawback detectors under a full evolutionary lifecycle, trained to agree with a ten-item anchored reference set, regularized by consensus over unlabeled outputs, and audited against a held-out anchor it never reads, yielding a transparent, inspectable metric rather than an opaque judge. Second, since no metric exists to beat, the yardstick is recovering what an accurate metric would have enabled, and \emph{Double Ratchet}, our co-evolution of the metric with a lifecycle-managed skill loop, does so: across code generation (MBPP+), enterprise text-to-SQL (Spider~2.0-Snow), and reference-free report generation, it retains 88--110\% of the held-out lift achieved by the same skill loop driven by ground truth or the best available rubric. Third, safety comes from anchor discipline plus outer audits: removing anchor guards collapses the metric into a vacuous detector while removing the lifecycle does not; and when evolved skills gamed the report rubric, an independent judge caught it, one detector repaired it, and a task-aware judge then preferred the evolved outputs over the pre-evolution baseline in 77\% of decided pairs. We argue this failure-expecting architecture is the right default wherever no reliable automatic verifier exists.

The paper introduces the Internet of Agentic Things (IoAT), an architectural framework that integrates agentic AI, IoT, cyber-physical systems, Physical AI, edge computing, and digital twins into a unified closed-loop orchestration framework. The proposed architecture consists of cloud, edge/fog, and physical IoT layers connected through autonomous AI agents that perceive, reason, coordinate, and actuate across distributed cyber-physical environments. The paper formalizes IoAT as a coupled workflow-control problem with nested strategic and tactical decision making using a hylomorphic dynamic programming framework that links agentic planning with physical execution. Smart-building orchestration is presented as a representative use case, and key research challenges related to safety, security, governance, resilience, and trustworthy deployment are discussed.

Multi-agent and memory-augmented LLM systems often place coordination content, shared state, prior discussion, tool outputs, summaries, and role instructions, inside the same finite prompt used for the current task. This creates a practical allocation problem: every token spent on coordination is unavailable to task instructions or evidence when a call is assembled under a fixed context budget. We introduce the Roundtable Context Window Test (RCWT), a controlled protocol for measuring this task-budget displacement effect. RCWT varies coordination content while controlling total budget, position order, task family, and scoring. In the main context-dependent recall task at $W=4096$, three commercial models remain near baseline through moderate overhead and then degrade sharply once residual reference evidence falls to a few hundred tokens. Window-scaling summaries are consistent with a task-specific residual-budget interpretation rather than a fixed percentage threshold, but we treat this as descriptive evidence rather than a universal law. To test whether the fixed-budget cliff persists when task evidence remains intact, we add an intact-task ablation: the full task/reference block is kept present while coordination tokens increase by expanding total prompt length. In that setting, all tested calls return every scored field correctly across GPT-4.1-mini, Claude Haiku 4.5, and Gemini 2.5 Flash up to a 95\% coordination ratio. This ablation narrows the claim: the main RCWT cliff is best read as task-budget displacement, not as proof that coordination volume alone causes semantic interference in the original open-ended task. RCWT is therefore a measurement primitive for context-allocation budgeting, not a complete theory of multi-agent benefit or session-level coordination.

Accurately assessing personality from text is challenging because traits are latent, context-dependent, and often subtly expressed across long narratives. Large language models (LLMs) offer new opportunities by processing extensive textual contexts, but pretraining of these models can induce latent "personality-like" biases, making single-model inferences inconsistent. We propose a fine-tuned multi-agent framework for detecting OCEAN personality traits, in which sub-agents are conditioned to adopt high, low, or neutral perspectives for each trait through masked language modeling (MLM) and psychometric supervision. A judge LLM aggregates and compares sub-agent outputs to generate final trait predictions, capturing multiple complementary perspectives while mitigating individual model biases. We evaluate the framework on life narrative dataset through quantitative and qualitative experiments, including baselines, ablations, and inference quality analyses. Our approach offers a scalable and interpretable method for text-based personality inference, highlighting the benefits of multi-agent reasoning grounded in psychometric supervision.