Relation
-
Agents-as-Peers: Bias-Controlled LLM Evaluation as Multi-Agent Peer Assessment
2026-10-20 07:13:23Yanki Margalit, Nurit Cohen-Inger, Erni Avram, Ran Taig, Seffi Cohen, Oded MargalitAbstract
Large language model (LLM) evaluation commonly relies on fixed human-authored benchmarks or centralized human/model judges. Fixed benchmarks can age and enter model training corpora, while centralized judges introduce their own scoring preferences. We cast LLM evaluation as a multi-agent peer-assessment problem and introduce Agents-as-Peers, a distributed evaluation protocol in which every agent generates tasks, answers the shared task set, and judges the answers of its peers, with no privileged evaluator. Self-ratings are excluded from the ranking, and judging is repeated under controlled identity and presentation regimes so that self bias, identity effects, presentation-position sensitivity, and judge severity are measured rather than left implicit. In an instantiation with 12 LLMs and 420 agent-generated questions, Agents-as-Peers aggregates 55,440 non-self peer judgments per regime. Most agents assign their own answers higher scores than those answers receive from peer judges; visible identity changes scores for some agents; the first fixed answer slot is associated with scores 0.39 points higher than the shuffled baseline; and judges differ markedly in severity. On labeled benchmarks, peer scores align strongly with accuracy on TruthfulQA (Pearson r=0.904) and GSM8K (r=0.873). On TruthfulQA, the peer-score correlation is substantially higher than the self-score correlation (r=0.538). A benchmark-overlap screen and item-level human audit further characterize the endogenous task set. These results support Agents-as-Peers as a cohort-relative measurement protocol whose ranking is reported together with the biases and rater effects that shape it. Code and dataset: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2602.02589
PDF
-
How Language Models Organize Task Information
2026-10-20 06:43:48Kaiyang Wan, Yuxia WangAbstract
Dense linear probes assess task readability from language model (LM) representations, but do not directly reveal how many features suffice for prediction. We measure readout sparsity with \(k^*\), the smallest feature count found by a fixed search protocol that recovers 95% of dense-probe AUROC above chance. We evaluate 19 LMs from five labs on 14 core tasks. Ten base models give the nine reliably readable tasks similar feature counts and nearly the same order. This agreement extends to the selected features themselves, which on three tasks show cross-model response correspondence beyond class-mean differences and a label-correlation-matched null. Instruction-tuned deployment selectively reduces \(k^*\), with both weights and input format contributing. In controlled synthetic experiments across 20 models, successfully learned targets admit single-feature readouts, whereas readable controls require tens of features. Reference-model \(k^*\) is associated with label requirements across eight students, following an empirical scaling of \(N^*\propto(k^*)^{0.65}\). After continual learning, saved heads can fail while linear readability remains high, and recovery requires fewer labels than initial task learning under the tested procedures. Pruning reduces linear readability, with wider pre-pruning readouts generally associated with larger losses. These findings distinguish task readability from readout sparsity and establish their complementary value for predicting learning requirements and diagnosing LM changes.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27724
PDF
-
Memory as Code: Declarative Retrieval over Codified Long-Term Memory
2026-10-20 06:17:51Aoting Zeng, Yuxin Du, Jianwei Wang, Yizhang He, Shan Jiang, Xuemin Lin, Kai WangAbstract
In long-term memory systems, memories are constructed from interaction histories and retrieved in response to queries, enhancing the reasoning ability of LLM agents. In complex scenarios, facts are scattered across sessions and their relationships evolve over time, making it difficult for existing memory systems to compose them to adapt to diverse queries. Motivated by the modularity and composability of code, we propose Memory as Code (MaC), a unified memory framework that jointly designs memory construction and retrieval through a shared code abstraction. MaC couples codified construction, which establishes executable memory abstractions, with declarative retrieval, which translates queries into structured retrieval programs. Specifically, codified construction builds a three-layer executable memory structure that progresses from an atomic function layer, through a relational layer, to a modular composition layer. Declarative retrieval recasts retrieval from relevance-based fact selection into declarative query processing over memory. Experiments on LoCoMo and LongMemEval show relative gains of up to 15.91% in overall LLM-judge accuracy over the strongest baseline in each setting, while substantially reducing token consumption, highlighting the effectiveness of memory as code for long-horizon memory tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27700
PDF
-
PanoGraphGS: Unifying Panoptic Segmentation and Hierarchical 3D Scene Graphs with Gaussian Splatting
2026-10-20 06:15:59Jinghan Gao, Siyu Wang, Dongyu Feng, Yiyuan Pan, Rong XiongAbstract
3D scene graphs support language-guided navigation by organizing physical entities and their containment and spatial relations into hierarchical maps. Building such structured maps requires a scene representation that consistently distinguishes semantic regions and object instances across viewpoints. To this end, we present PanoGraphGS, a unified framework that integrates panoptic 3D Gaussian representation learning with hierarchical 3D scene graph construction. The framework learns semantic and instance features on shared Gaussian geometry. Prototype-based semantic learning and association-based instance learning produce consistent panoptic predictions across viewpoints. The learned representation supports novel-view panoptic segmentation and direct Gaussian-level panoptic decoding. A semantic-aware scene graph module further recovers a floor-room-stuff-object hierarchy with containment and spatial relations. We further release NaviGS, a real-world dataset for panoramic scene understanding. The dataset provides 3DGS, mesh, RGB, semantic, and depth data. For 3D scene graph identification, PanoGraphGS outperforms the strongest evaluated baseline in object-node accuracy by more than 50\% points on average across Replica and NaviGS. PanoGraphGS also achieves state-of-the-art performance among the evaluated methods in 2D panoptic segmentation. These results demonstrate the effectiveness of PanoGraphGS in both hierarchical 3D scene understanding and 2D panoptic segmentation. The project website is available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27696
PDF
-
Inferring continuous cognitive traits from limited neuroimaging data using pairwise convolutional neural networks
2026-10-20 05:32:22Marjan Rashidi, Robert Bain, Craig Stark, Jeffrey Krichmar, Mary Hegarty, Elizabeth ChrastilAbstract
Individual differences in cognitive abilities vary widely, yet the structural brain features supporting these differences remain poorly understood. Traditional brain-behavior analyses rely on predefined metrics such as volume or cortical thickness and typically assume linear relationships with behavior, limiting sensitivity in small datasets. Here, we introduce a pairwise convolutional ranking framework that learns directly from whole-brain T1-weighted MRI by making comparative rather than absolute predictions, which performs well even with small samples (N=108). We tested this method with four small-scale spatial abilities tasks. Regression analysis predicted performance with high accuracy across all tasks, achieving greater predictive power over standard regression and pointwise deep learning baselines. Iterative cortical-region removal (backward elimination) revealed brain regions supporting individual performance common to all four tasks, as well as task-specific regions. These results demonstrate that pairwise learning extracts meaningful behavioral information from structural MRI, providing prediction and interpretable insight into the neuroanatomy of individual differences.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27668
PDF
-
TriGeoBench: A Multimodal Mathematical Geometry Benchmark with Figures and TikZ
2026-10-20 04:52:30Jing Wang, Yitong Yang, Hao Dou, Yaomin Wu, Shuting He, Tingfei Shen, Jia Wu, Jiaqi Li, Yinglin WangAbstract
Multimodal Large Language Models show strong vision-language capabilities, yet their geometry-centric reasoning remains insufficiently understood. Existing mathematical benchmarks typically rely on rasterized figures and cover limited geometric knowledge areas and response formats, making it difficult to assess whether models reason over geometric structures or exploit superficial visual patterns. We introduce TriGeoBench, a geometry-centric multimodal benchmark constructed from Chinese middle- and high-school mathematics examinations. TriGeoBench contains 7,000 question-answer pairs and 12,235 geometric figures with TikZ code, spanning plane geometry, solid geometry, analytic geometry, function, and vector geometry across diverse response formats. Each instance forms a {question-answer pair, image, TikZ} triplet: question-answer pairs are stored in LaTeX, images as RGB figures, and TikZ preserves geometric topology, symbolic annotations, and relational structure. We evaluate representative open- and closed-source MLLMs through image-to-TikZ and multimodal reasoning probes. Results reveal persistent fine-grained structure and image-text alignment challenges. Oracle experiments with human-curated reference TikZ suggest that accurate executable representations provide complementary geometric cues; they do not evaluate an end-to-end conversion pipeline. TriGeoBench offers a testbed for structure-aware mathematical reasoning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27626
PDF
-
HumanSkill-4D: Training-Free Language Querying of Animatable Gaussian Avatars in Space and Time
2026-10-20 04:37:26Junsu Kim, Naeun Kim, Dongyoon Han, Seungryul BaekAbstract
We introduce avatar referring segmentation, which selects regions of animatable Gaussian avatars from language queries. A query can also specify a pose or motion to select matching intervals. This lets users interact with avatars for local editing and motion inspection. However, making these selections with existing language Gaussian splatting methods poses two challenges: adjacent parts and garments are hard to separate and may be referred to indirectly (region ambiguity), while fine-grained motions can look alike in RGB (motion ambiguity). We address these challenges with HumanSkill-4D, a training-free framework that shares a multimodal large language model (MLLM) and geometric operators across avatars. For each avatar, a semantic memory links body-part and garment categories to persistent Gaussian regions and appearance descriptions. The MLLM uses category names and descriptions to select categories by appearance, relation, or function. To find matching intervals, it composes operators into programs that check query conditions on driving joint trajectories stored in a motion memory. These category choices and operator outputs make selections traceable, while composing or adding operators supports new motion conditions. To evaluate this task, we introduce three human-annotated benchmarks and show that HumanSkill-4D outperforms the baselines in part segmentation and 3D and 4D referring segmentation, with compact memories and only minutes of setup per avatar.
Abstract (translated)
URL
https://arxiv.org/abs/2610.24452
PDF
-
Local Allow, Global Deny: Why Per-Action Authorization Does Not Compose in LLM Agents
2026-10-20 04:35:39Anmol Pandey, Aditya Jain, Liang Chen, Carsten MapleAbstract
Autonomous LLM agents act through tools under per-action authorization: a local guard screens each tool call in isolation. We show that local authorization does not compose into global authorization. A compositional authorization failure (CAF) is a locally authorized execution, or a pair of locally authorized executions, that violates a global policy spanning intent, purpose, provenance, consent, and delegation. We characterize CAFs along two axes, expressiveness (single-trace safety versus relational k-safety) and observability (a ladder of monitor profiles). Our central result is an impossibility: relational purpose violations form a 2-safety hyperproperty with no per-trace locus of blame, so a per-action monitor that never raises a false alarm detects none of them, independent of the underlying model or decision procedure; they become visible only under cross-execution monitoring, which none of the six guardrails we survey implements. We make this constructive with AL-SYNTH, a SAT/SMT synthesizer (using no language model) that emits minimal CAF witnesses, each checked against an exhaustive oracle. The empirical signature is sharp: across nine LLM judges the per-action view approves the synthesized witness almost without exception, blocking it at a rate comparable to a benign control, yet the same models flag it when shown the two executions together. Per-action authorization is therefore structurally insufficient for relational policy conjuncts not reducible to single-trace safety: to bound them, a monitor must compare executions or accept false alarms.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27620
PDF
-
JARVIS: From Visual Manuals to Autonomous Robotic Assembly
2026-10-20 04:32:37Fan Ding, Fucai Ke, Kevin Yuchen Ma, Xuewen Luo, Wenyu Liang, Vishnu Monn Baskaran, Hamid Rezatofighi, Yan WuAbstract
Visual manuals communicate procedural knowledge through stepwise diagrams, but translating these instructions into reliable long-horizon robotic execution remains challenging. We introduce JARVIS, a method-agnostic benchmark for autonomous visual-manual-guided execution without additional language instructions. JARVIS comprises 236,679 LEGO assembly tasks across 8,987 scenes, with a carefully curated JARVIS-v1 suite of 50 tasks for interactive robotic evaluation. Tasks span varying initial structural complexities and assembly horizons, requiring agents to interpret manuals, ground actions in an evolving workspace, and execute the intended procedure. A structure-aware evaluation protocol captures geometric accuracy, structural relations, construction order, partial progress, and completion. We further introduce JARVIS-agentic, a tool-grounded embodied agent that composes robot primitives into executable programs and uses execution feedback for verification and recovery, without task-specific action trajectories or model parameter updates. Across nine multimodal and reasoning models, task-oriented primitives increase the average assembly score. Performance nevertheless declines sharply with assembly horizon, and execution errors remain difficult to correct. Prior-task experience improves initial performance, while repeated reflection provides inconsistent benefits, highlighting persistent limitations in connecting visual procedural reasoning with reliable physical execution.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27618
PDF
-
HiQGS: Hierarchical Spatiotemporal Querying for Feed-Forward Gaussian Reconstruction
2026-10-20 04:10:57Qiaowei Miao, Qifeng Cheng, Yijing Guo, Yuru Xiao, Ziqian Ni, Sheng YangAbstract
Feed-forward Gaussian reconstruction from sparse posed monocular frames requires reasoning over observations that vary jointly in viewpoint and time. Existing methods either predict Gaussian primitives from observation-aligned features or aggregate multi-frame information through compact latent queries. However, the former tightly couples Gaussian allocation to individual frames, while the latter does not explicitly structure observed evidence according to its temporal relation to the requested timestamp. Consequently, evidence across spatial scales, viewpoints, and timestamps can be difficult to align and integrate. We present \textbf{HiQGS}, a query-based hierarchical framework for feed-forward Gaussian scene reconstruction from sparse monocular frames. HiQGS introduces a \textbf{Hierarchical Spatiotemporal Encoder} that organizes multi-resolution geometry-appearance visual features according to image location, observation time, and camera pose, thereby preserving complementary appearance and structural cues. Our \textbf{Time-Adaptive Decoder} use target timestamp queries the resulting hierarchical representation to retrieve relevant spatiotemporal evidence and predicts the corresponding Gaussian scene state. Experiments on static and dynamic scene reconstruction benchmarks demonstrate strong rendering quality and geometric accuracy, validating the effectiveness of hierarchical visual encoding and target-time-aware Gaussian querying.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27604
PDF
-
WORTH Benchmark: Evidence Discovery under Not-Fully-Cooperative Information -- A Two-Episode Pilot
Abstract
Existing reasoning evaluations typically provide task-relevant information to models as available premises, rarely examining concurrently how models discover evidence autonomously and how information presentation within narratives guides or obscures judgment. This paper proposes the WORTH Benchmark: models read original mystery narratives without clue annotations, autonomously select key clue sentences, identify case participants, and explain their reasoning process. The narrative materials are designed by professional authors such that evidence facilitating reasoning coexists with narrative arrangements that delay the revelation of truth. We report a pilot experiment under full-story conditions across two episodes, conducting a total of 70 formal runs on models, of which 67 runs are content-scorable. Scoring separately records hits on author-designated key clue sentences and participant selections; natural language explanations are used for case diagnosis and are not scored independently. The results show that clue hits and participant judgments are not synchronized; across 37 content-scorable runs in Episode 02, 31 runs simultaneously selected the same set of incorrect participants, while adjacent suspicion-oriented and exoneration-oriented sentences were selected 25 times and 0 times, respectively. These phenomena illustrate the diagnostic value of component-wise analysis and suggest that not-fully-cooperative informational relationships may influence evidence selection and judgment. This study only implements full-story conditions for two episodes, and does not establish stable model rankings or infer general mechanisms on this basis; human player records have not yet been cleaned, and this version does not report human-model comparisons.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27597
PDF
-
Factual Access States in Diffusion Language Models: From Mechanism to Knowledge Editing and Retrieval
2026-10-20 03:45:45Yuchen Guo, Zhuoyuan Mao, Yiwen GuoAbstract
Factual recall is a lookup: some representation inside the network serves as the key, and what that key is made of decides which queries can retrieve the fact, what an edit must cover, and what a retriever can read. We call it the factual access state. In autoregressive (AR) models the subject position enriches attributes largely independently of the relation, and relation-specific selection happens later, at the readout position. In masked diffusion language models (DLMs) we find the state is assembled differently, and much earlier: in the first attention layer the wording of the relation is routed into the position of whichever entity is visible and the early feed-forward layers look the fact up on this fused representation. The state therefore varies with the relation's surface form and with the query direction, so the queries about one fact occupy a region of state space rather than a point, and an operation on the fact succeeds only for queries that enter it. We treat the state as an interface to a DLM's factual knowledge, and put it to two uses. Writing to the state gives a knowledge editor: from one declarative statement, Multi-View Edit fits one shared value per direction across that region and installs it in the early-MLP block, preserving source-prompt reliability while substantially improving paraphrase and reverse question answering, where AR editors gain near zero. Reading it suggests a complementary use: the state at each entity position can serve as a relation-specific retrieval vector with no training, returning both pieces of evidence for two-entity queries on which pooled, prompted and lexical retrievers almost never return both.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27580
PDF
-
Multiscale Temporal-Spectral-Spatial EEG Learning With Modality-Specific and Modality-Invariant Representations for Emotion Recognition
2026-10-20 02:35:16Jian Zhang, Manrong Liu, Yilei Huang, Wenhua Liu, Xiaohui Yang, Lijun YangAbstract
Multimodal physiological signals provide complementary cues for emotion recognition. They also raise two related challenges. Emotion-related EEG patterns vary across time, frequency bands, electrodes, and brain regions. Different physiological modalities also contain both modality-specific and shared affective information. This paper presents GMSI-Net, an EEG-centered multimodal emotion recognition framework. It has three key components. The temporal-spatial-spectral-region (TSSR) encoder models multiscale temporal-channel patterns, differential entropy features, and local-to-global brain-region relations. The modality-specific representation (MSR) branch uses bidirectional cross-modal attention and learnable gates. It controls information transfer while retaining modality-specific cues. The modality-invariant representation (MIR) branch learns a shared affective representation. An adaptively weighted contrastive loss aligns this representation across modalities. Independent multiscale Conv-Transformer encoders process the peripheral physiological signals. GMSI-Net is evaluated on DEAP and MAHNOB-HCI using within-subject random ten-fold cross-validation. The model reaches 96.21 (1.47) and 96.41 (1.60) for valence and arousal on DEAP. It reaches 94.10 (2.86) and 94.48 (3.77) on MAHNOB-HCI. These results indicate that combining EEG features with complementary physiological information can improve emotion recognition.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27521
PDF
-
Cross-Visibility-Gap Identity Persistence for 4D Scene Graphs
2026-10-20 02:20:34Jiahui Hu, Li Gao, Yu Liu, Pengzhan ZhouAbstract
4D scene graphs require persistent object identity when an object disappears inside an opaque carrier and later reappears as a fresh detection. When appearance and target-centric motion are unavailable, identity recovery must rely on relational history. We formulate cross-visibility-gap identity persistence as historical-node recovery through a remembered target--carrier binding. We introduce CVG-SGM, a perception-frontend-compatible transfer--maintain--read memory. Geometry-weighted cross-node interaction transfers target-specific evidence into carrier states, a shared per-node GRU maintains this information during transport, and a two-time readout combines stored carrier--identity compatibility with current detection--carrier geometry to recover the historical identity. In the controlled one-identity-per-carrier setting, CVG-SGM achieves 1.000 re-link accuracy, compared with 0.570 for both no-memory and parameter-count-matched no-write controls. It also generalizes to scene-disjoint, unseen longer visibility gaps. On real RGB-D/SAM2 feature-level episodes, it correctly re-links 36/36 held-out detections, compared with 28/36 for the strongest fixed carrier--identity control. Counterfactual carrier-memory swaps redirect predictions according to the swapped bindings, supporting the role of stored relational information in identity recovery. The current approach remains limited by shared-carrier identity ambiguity and does not provide end-to-end raw-video re-identification.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27505
PDF
-
MiCLA: Mixture-to-Constituent Latent Alignment for Polyphonic Audio Representation Learning
2026-10-20 02:10:10Kohei Yamamoto, Kosuke OkusaAbstract
Real-world audio typically contains concurrent sound events, yet standard self-supervised methods treat each clip as an isolated training unit without explicitly modeling mixture--constituent relationships. We propose Mixture-to-Constituent Latent Alignment (MiCLA), a two-stage self-supervised pre-training framework that uses synthetic log-mel mixing to align mixture representations with their constituent clips. Stage 1 structures single-clip representations via masked self-distillation and contrastive regularization. Stage 2 introduces bidirectional alignment: for a synthetic mixture, MiCLA suppresses the component aligned with one constituent and aligns the residual with the other, using projective residualization on patch tokens and sample-adaptive soft residualization on the global class token. In downstream evaluations, MiCLA achieves 43.2% mean average precision (mAP) on AS-20K and 50.3% mAP on AS-2M under fine-tuning, and achieves the highest mAP among evaluated models on both AudioSet settings under frozen Convex Gated Probing (CGP). Furthermore, a Joint Semantic Retrieval diagnostic shows that MiCLA improves simultaneous retrieval of both constituent classes over target-averaging baselines. These results indicate that directional mixture-to-constituent alignment provides an effective pre-training signal for polyphonic audio representation learning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27491
PDF
-
How Data Scales in Agentic Reinforcement Learning: Laws and Synthesis Strategies
2026-10-20 01:59:17Bowei He, Yankai Chen, Xiaokun Zhang, Changjiang Han, Ye Yuan, Chong Li, Xue LiuAbstract
Pretraining scaling laws treat training data as a one-dimensional quantity: a token count. Agentic reinforcement learning (RL) inherits this scalar abstraction, but modern synthesis pipelines now generate \textit{tasks}, \textit{environments}, and \textit{trajectories} independently and at very different costs, turning data into a structured design space. This raises a question pretraining never had to answer: given a synthesis budget, \textit{which axis} should be scaled? We address this question end-to-end with three tightly coupled contributions. First, we decompose agentic data into three cardinality axes (task, environment, trajectory), plus a reward-density control variable, and fit per-axis scaling relations on different large language models (LLMs) across math+code, SWE, web, and UI/tool application domains. Across all model sizes, model families, domains, and RL algorithms (PPO and GRPO) we measured, the scaling exponents order as environment $>$ task $>$ trajectory. When environment diversity is decomposed into world and environment-structure diversity, its advantage is mainly driven by the former, while the latter is just comparable to the task diversity. Second, we use the fitted laws as a common ruler to price synthesis methods, capturing each method's efficiency as a per-axis discount factor and identifying critical synthetic-to-real ratios at which collapse begins; these ratios differ by an order of magnitude across axes. Third, we cast budget allocation as a constrained optimization under our laws and discounts, deriving a Chinchilla-style compute-optimal recipe and validating it at held-out scale. The empirical results demonstrate that following the recipe yields up to $1.7\times$ compute efficiency over the dominant ``scale-the-trajectories'' practice.
Abstract (translated)
URL
https://arxiv.org/abs/2610.23946
PDF
-
Trail-Weighted Influence Scores for Approximate Outcome Prediction in Bayesian Networks
2026-10-19 22:46:24Sahil Kumar Rajput, Randhir KumarAbstract
We study trail-weighted pruning of evidence for Bayesian-network prediction. A traversal enumerates the active trails from each evidence node to the outcome, scores each trail by the product of its edge strengths, and drops evidence whose score falls below a threshold. For collider-free networks (chains and forks) over binary variables, we show that the trail score equals the exact correlation between a trail's endpoints, that a chain and a fork with equal edge strengths have identical joint distributions, and that the pruned posterior is within a computable sum of per-trail terms of the exact posterior when each dropped trail lies in its own branch of the outcome node. We prove this bound, check it exhaustively on all small trees, and show that it fails when dropped trails share a branch. In simulations on random trees, thresholding trail scores gives 1.4 to 13 times lower error than random pruning of the same size. We also apply the score, as an unproven heuristic, to seven bnlearn networks with colliders and loops. Keeping the top half of the active evidence by trail score gives lower error than random pruning on all seven, but it is not clearly better than a mutual-information ranking, and scoring costs more than the inference it saves on six of the seven, so we show no speedup. Models trained on trail, graph and information features and tested on networks they never saw do not beat the trail score or mutual information, but trail features carry most of what they use and clearly outperform generic graph features. A bound for general networks is left to future work.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27366
PDF
-
Predictive Closure: A Set-Level Characterization of JEPA's Feature Selection
2026-10-19 22:43:59Mosh Levy, Aly Lidayan, Yoav GoldbergAbstract
In joint-embedding predictive architectures (JEPA), an encoder and a predictor are trained together to produce a representation that minimizes the error in predicting the representation of the next observation. Because the representations of both the current and the next observation are learned, it is unclear which features they will include. Earlier characterizations describe the features JEPA selects in terms of their individual predictability, raising the concern that unpredictable features, like a traffic light, will not be encoded into the representation. However, the prediction error is also affected by the relationships between features, e.g., the light's color helps predict whether the traffic will stop. We propose the predictive closure characterization, which accounts for these relationships. Based on a reformulation of the JEPA training objective, we show that selection depends on properties of feature sets. This view agrees with earlier works when features evolve independently, but makes new predictions when they do not, for instance that a feature like the traffic light can still be retained. We test this characterization in controlled experiments, which show that unpredictable features can become better represented when they help predict other features, and that changing one feature's predictive relationships can affect how well others are represented. We also demonstrate implications for steering feature selection: targeting one feature with an auxiliary objective, or supplying it to the predictor as side information, can strengthen the representation of another feature whose future values it helps predict, so steering can be indirect.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27364
PDF
-
On the LLM Verification for Argument Graph Generation
2026-10-19 22:41:41Jakub B\k{a}ba, Jaros{\l}aw A. ChudziakAbstract
Multi-agent LLM systems can formalize natural-language argumentation end-to-end, turning plain text into structured graphs by classifying components, connecting them with relations, and enriching the resulting graphs into formal models. Quality assurance for those graphs is almost always another LLMs: a reviewer, judge, or critic agent. On the other side, in adjacent neuro-symbolic fields, such generated output is often verified symbolically by external checks and solvers. In this work, we argue that in argument graph verification, the allocation should be similar, and based on what each instrument can guarantee. We propose layered verification: starting from deterministic checks, moving to solver-based and LLM reviews, and utilizing debate only under uncertainty, leaving human judgment for what no other instrument can decide. We ground our principle with a study on human-annotated graphs with injected structural defects. An LLM reviewer detects them at near-ceiling recall, including cycles in the support relation, but it objects to almost one in a ten correct graphs, even though it should not, happens to be inconsistent in verdict on identical inputs and is significantly slower than the equivalent exact check. It shows that while LLM reviewer does not lack the ability to detect structural defects, it is worth having a deterministic verification gates.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27361
PDF
-
Vulnerability Analysis for Safe Multi-Agent Reinforcement
2026-10-19 21:09:02Taozheng Zhu, Shixiong Jiang, Fanxin KongAbstract
Safe Multi-Agent Reinforcement Learning (SMARL) aims to learn cooperative control policies under prescribed safety constraints. Al- though existing SMARL methods focus on enforcing safety during policy optimization, their vulnerability remains largely unexplored. This paper investigates inference-time observation attacks against SMARL, where an adversary perturbs the observations of a subset of agents, thereby altering the actions produced by their policies. We propose MinMax Attack, an attack framework that jointly selects adversarial target actions for compromised agents while accounting for the Cooperative Resistance of benign agents. We further derive a bound on the induced safety cost and characterize its relation- ship with the Compromised-Agent Budget and the Perturbation Budget. Experiments across multiple safety-critical multi-agent benchmarks and representative state-of-the-art SMARL algorithms show that MinMax Attack induces substantially higher safety costs than existing attack baselines while preserving task reward, thereby reducing its detectability and achieving both attack effectiveness and stealthiness.
Abstract (translated)
URL
https://arxiv.org/abs/2610.27293
PDF
