跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+816新增论文
+642其中 AI 相关
+95已有论文修订
+95新增 AI 导读

arXiv 最近完成扫描:2026-11-21 12:19 · 全站最近入库:2026-11-21 12:19各信源进度 →

统计口径与数据新鲜度

统计 2026-11-20 13:14 — 2026-11-21 13:14(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-21 12:19最近发现新增版本:2026-11-21 12:19库内论文最新更新时间:2026-11-20 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

29,158 已收录论文24,181 AI 相关8,992 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 6,976 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-21 12:17(北京时间)
新论文 · v1arXiv:2611.23648

Toaster: Nonlinear Memory from Cascaded Linear Memories and Target Propagation

Zeman Li, Ali Behrouz, Yuan Deng, Praneeth Kacham, Rami Al-Rfou, Meisam Razaviyayn, Vahab Mirrokni

原文摘要

Recurrent memories let language models process long sequences with a fixed-size state. Linear memories such as DeltaNet and KDA store one matrix and take one delta-rule step per token. The product of a chunk of such steps has an exact WY form, so they train fast. Deep memories such as TTT-MLP and Titans store a small MLP, which can hold maps that no matrix can, but their update has no WY form. Toaster bridges the two by building on a known fact: a gradient step on a linear layer is a delta-rule write. It builds a two-layer nonlinear memory from two linear memories with an activation between them. We use target propagation to give the hidden memory its values by projecting the output error back through the second memory. Given a chunk's cross-layer signals, approximated once per chunk, each layer can then run on an unmodified delta-rule kernel. Our default runs this kernel on the output layer and updates the hidden layer once per chunk. Running it on both layers gives a better model in our ablation. At matched backbones and training-token budgets, Toaster has lower loss than KDA at all five tested sizes (0.2B-1.6B non-embedding parameters) on both C4 and PG19. Its C4 perplexity also matches or slightly improves on the attention baseline at every size. At 1.2B it is also below Titans. In a 3:1 hybrid with global attention, the Toaster hybrid has the lowest perplexity of the four hybrids tested, and it is 0.18 (C4) and 0.44 (PG19) below the attention baseline. At the largest size, Toaster's mean three-shot accuracy on 13 tasks is 51.84%, compared with 50.52% for KDA and 50.04% for Titans. In length extrapolation, Toaster's loss stays below KDA's up to 20K tokens (5x the training length). On the 145M backbone, at a hidden width matched to Titans' memory, Toaster takes at most 1.33x KDA's training-step time and is 2.8-5.5x faster than the tested Titans configurations at 32K context.

cs.LG · cs.AI

版本更新 · v2arXiv:2610.02527

CriticHack: Evaluating Visual Rewards Under Robot Policy Optimization

Jiaxuan Luo, Xingguo Xu, Shanshan Wang, Yuhan Zhou, Zhen Zhang

原文摘要

Learned visual reward models score a robot's execution from camera images and an instruction, and are increasingly used to fine-tune robot policies. Two checks decide whether to trust one: the reward model separates successes from failures, and task success rises during training. We show that both checks can pass while the policy learns to act on the wrong object. When a diffusion policy is fine-tuned against the reward model Robometer to open a drawer, task success rises by 10.2 percentage points. Wrong-object failures, which open a different drawer, rise by 10.9 points, whereas the same training on the simulator's task-completion signal raises success alone. A second reward model, RoboReward-8B, behaves the same way, and the effect reappears on a physical robot. This happens because of how the reward scores the policy's own failures. Most failures make little progress and score low, while a wrong-object failure looks like progress and scores well above them. Optimization shifts a policy toward outcomes that score above the executions they compete with, the executions they can take the place of. It therefore turns low-scoring failures into both successes and wrong-object failures, even when the reward model scores successes higher. This rule needs only executions of the initial policy, so it can be tested before training. On a stacking task we predicted which of three object placements Robometer would exploit, and training raised wrong-object stacks from 20% to 69% in that placement and in neither of the others. The same rule gives a check to run before optimization and a repair that needs no labels: scoring each execution again on frames re-rendered with the objects the instruction does not name put back removes the amplification on both tasks and on the robot.

cs.RO · cs.AI · cs.LG

新论文 · v1arXiv:2611.23639

Searching the Pareto Front: Multi-Objective Retrieval for Heterogeneous Scientific Modalities

Song Young Oh, Nesar Ramachandra, Bin Xia, Ian Foster, Kyle Chard

原文摘要

In cosmology, surveys span different sky regions and wavelength ranges, resulting in uneven modality coverage across observational datasets. Such heterogeneity can be addressed through either multimodal foundation models (FMs), which learn shared embedding spaces, or independently learned modality embeddings combined via late fusion. We investigate whether the latter provides an adaptive alternative to FM-based scientific retrieval. We evaluate this approach on a galaxy retrieval task spanning diverse modalities--history of star formation, spectra of light emitted, and multi-wavelength observations from two telescopes--comparing an FM baseline with separately trained embeddings fused through multi-objective optimization based on Pareto optimality. We find that the FM baseline achieves strong embedding-space proximity, but our method preserves modality-specific distance structure that allows scientists to inspect which physical signals drive similarity. Our method also achieves improved fidelity in raw-data space: across six measured properties, its nearest neighbor is 8% closer to the query on average than that retrieved by kNN in the FM embedding space (spread 0.128 vs. 0.139). These results suggest that multi-objective retrieval can surface astrophysically coherent candidates that single-objective search alone may overlook.

cs.AI · astro-ph.IM

首次观测 · 记录 #1软件学报:jos:7749软件学报

面向大语言模型智能体安全的异常检测方法综述

面向大语言模型智能体安全的异常检测方法综述

马云聪, 宋泓憬, 黄辰林, 于丰源, 程华, 丁滟, 杨任宇, 李爱平 等 9 位作者

AI 速读 基于摘要

该综述从模型层、认知层和交互层梳理LLM智能体安全威胁,提出基于双层复合图的异常检测方法分类,并构建多维度评估指标。

cs.* · cs.CL · cs.AI · cs.MA

软件学报 · 2026 · 正式发表

查看原文摘要

以大语言模型(large language model, LLM)为核心的智能体和多智能体系统, 凭借其卓越的意图理解、推理规划和执行能力, 正广泛应用于各个领域. 近年来, 随着LLM能力的提升及其与外部环境交互的加强, 针对智能体系统的攻击行为持续演化, LLM幻觉也难以避免, 导致智能体系统在实际应用中面临各种安全威胁. 现有研究表明, 仅依赖LLM自身的安全机制难以保证系统的安全性, 准确识别各种异常行为对于智能体生态的良性发展有重要意义. 围绕智能体安全的异常检测问题, 首先从模型层、认知层和交互层梳理了当前智能体系统面临的安全威胁; 其次, 从拓扑图视角对智能体和多智能体系统进行形式化表示, 提出基于双层复合图的分类方法, 对LLM智能体安全的异常检测方法进行综述和分析; 接着, 提出提升智能体可信度的6大安全目标, 并基于此构建一个多维度的异常检测评估指标, 且在主流模型下验证其有效性和鲁棒性; 最后, 讨论现有检测机制的主要挑战和未来发展方向.

新论文 · v1arXiv:2611.23631

FairPGT: Partitioned Graph Transformers for Dual-Level Fair Graph Clustering

Siamak Ghodsi, Amjad Seyedi, Sahar Tahmasebi, Eirini Ntoutsi, Sören Auer

原文摘要

Clustering on attributed graphs can produce demographically biased clusters, particularly when communities align with protected groups. Graph transformers can amplify this bias through same-group attention, while global attention limits scalability. We propose FairPGT, a partitioned graph transformer for fair graph clustering. FairPGT restricts attention to METIS partitions and connects them through routing tokens, which divides the quadratic cost of attention by the number of partitions. A graph-aware self-training target guides clustering, with a provably bounded deviation from the standard DEC target. Fairness is imposed at two levels: a mutual-information term promotes demographic parity in cluster assignments, while a normalized cross-group attention ratio (nCGAR) guides attention toward group neutrality through self-distillation. We prove that the nCGAR correction moves attention toward neutrality and establish a condition preventing overshooting. Across seven benchmarks, FairPGT achieves the best clustering performance and statistical parity on the three largest graphs (Credit, Pokec-n, Pokec-z), with a statistical-parity gap approximately ten times lower than DMoN, and is Pareto-dominated on only one dataset. On controlled stochastic block models, FairPGT attains 97% of the derived upper bound on NMI for exactly fair clustering when communities and protected groups coincide. These graphs further show that, under strong alignment, assignment-level fairness alone makes the attention group-heterophilous, whereas attention-level fairness alone improves neutrality without ensuring fair clusters. Code for reproducibility is available at https://github.com/SiamakGhodsi/FairPGT.

cs.LG · cs.AI

新论文 · v1arXiv:2611.23629

The Value Semantics Layer: Enabling Machine-Understandable Databases for Agentic AI

Jalal Mahmud, Eser Kandogan

原文摘要

As data systems increasingly support autonomous LLM-powered agents alongside human analysts, they require richer semantic infrastructure bridging analytical concepts and analytical operations. Existing metadata systems expose schemas, catalogs, lineage, and governance information, but provide limited support for representing operational semantics. We propose the Value Semantics Layer (VSL), an architectural abstraction that externalizes value semantics as persistent, planner-visible infrastructure. VSL introduces Value Semantic Types (VSTs), reusable operational semantic assertions capturing distributional, relational, process, quantitative, and policy-grounded semantics. Unlike descriptive metadata, VSTs enable planners to validate analytical operations, select appropriate operators, and guide analytical workflows. We present the motivation, formalization, architecture, and prototype realizations of VSL, and outline a research agenda toward AI-native semantic infrastructure for analytical systems.

cs.DB · cs.AI

版本更新 · v2arXiv:2608.06400

Beyond Routing Weights: Faithful Response-Level Interpretation of Mixture-of-Experts Reward Models via Contribution Contrast

Yifan Wang, Jinyi Mu, Mayank Jobanputra, Yu Wang, Soyoung Oh, Isabel Valera, Vera Demberg

原文摘要

Reward models are central to learning from human preferences, yet identifying what drives their predictions remains challenging. Recent sparse Mixture-of-Experts (MoE) reward models seek to improve interpretability by routing prompts to specialized experts and characterizing experts through examples with high routing weights. However, routing weights only reveal which prompts an expert $\textit{receives}$, not how it $\textit{judges}$ responses, providing only a partial account of expert behavior. We therefore propose $\textbf{Co}$ntribution-$\textbf{Co}$ntrast ($\textbf{CoCo}$) response-level interpretation, which faithfully characterizes experts' roles using chosen-rejected response pairs with the largest contribution contrasts, jointly capturing routing and preference behavior. Across automatic and human evaluations, CoCo yields more coherent, faithful, and specialized interpretations than router-based, score-based, and sparse autoencoder-based alternatives while maintaining competitive reward modeling accuracy. Overall, we provide a controlled study of how different example-selection signals affect the interpretability of MoE reward models.

cs.AI

版本更新 · v5arXiv:2512.20974

Meta-RL with Bayesian Linear Task Models

Jingyang You, Hanna Kurniawati

原文摘要

Deep Bayesian reinforcement learning adapts to unseen tasks by inferring latent transition and reward models, but existing methods typically rely on variational posteriors and evidence lower bounds, introducing approximation error and unstable task representations. We introduce GLiBRL, a deep Bayesian RL framework that combines generalised linear task models with learnable non-linear basis functions. GLiBRL features conjugate Bayesian inference, yielding exact, sequential posterior updates over task parameters and model noise, together with a closed-form marginal likelihood that eliminates variational inference. The update is naturally permutation-invariant, allowing GLiBRL to integrate with both off- and on-policy algorithms. GLiBRL also learns task representation admitting an exact kernel identity, relating distances between task representations to kernel discrepancies over the task contexts. Compared against eight representative or recent meta reinforcement learning methods, GLiBRL achieves the highest aggregate zero-shot test performance on both the MuJoCo locomotion and MetaWorld manipulation benchmarks.

cs.LG · cs.AI · cs.RO

新论文 · v1arXiv:2611.23613

MANGO Arena: Same Seeds, Different Chains Evaluating LLM Agents in a Disrupted Supply Chain

Sara Amar

原文摘要

We introduce MANGO Arena, an open-source testbed in which four LLM agents each run one company in a serial supply chain adapted from MANGO, a classroom factory-management game. The environment enforces hard rules (a truck limit, a no-credit cash constraint and production capacity), logs every blocked attempt to break them, and gives agents a public message board on which side payments are explicitly forbidden. This lets one environment measure three things at once: an economic outcome (the bullwhip effect), rule adherence under pressure, and proposals to collude. In a pilot with three small reasoning models, two scripted reference policies, two conditions (with and without a supplier fire) and three demand seeds, all LLM agents amplified demand variability as it moved upstream, by 7 to 25 times at the Retailer and 13 to 179 times at the Supplier, often more than a textbook order-up-to policy. They did so largely by ordering in batches (zero orders in up to 60% of decisions). Real rule violations were rare (at most 7 of 288 decisions per model) and no agent proposed a money side deal. The pilot's main lesson concerns evaluation method. API failures (rate limits and runaway hidden reasoning) entered the metrics silently as "no order" decisions. And when we reran games with the same seeds and no failed calls, upstream bullwhip changed by up to 17 times between runs. With three seeds, conclusions about how a disruption affects a given model did not survive a rerun. We report these failures in full and propose reporting practices for multi-agent evaluations of this kind. Code, prompts and every logged decision are public.

cs.MA · cs.AI

新论文 · v1arXiv:2611.23611

RefinementGuard: Counterexample-Directed Semantic Qualification for Hidden Agent-Tool Execution

Adhyayan Veer Singh, Fanjiang Ye, Yuke Wang

原文摘要

Tool runtimes can lower an agent-selected operation into a hidden parallel graph, reducing latency and context while risking semantic mistranslation. We introduce RefinementGuard: translation validation for hidden, effectful agent-tool lowerings, coupled to artifact-bound, fail-closed runtime admission. Qualification compares a candidate graph with an independent reference on isolated state using explicit parent-operation obligations. We evaluate two repository-level contracts, bug localization and patch validation, on controlled repositories. Among 100 activated, non-equivalent development/qualification faults adjudicated by one human, parent-semantic validation rejects 49/100 (operator-cluster bootstrap 95% CI [33.7%, 64.8%]). The objective is selective admission, not maximum mutation rejection: exact equality rejects 100/100 faults but also all 60 valid variants, whereas obligation-based validation accepts 60/60; its 51 residual misses identify concrete specification gaps. On a frozen disjoint input split, the held-out B2-B3 result is localization-only because 21 patch mutants fail structurally; B3 rejects 38/56 localization defects. In a controlled asynchronous-delay regime, the localization serial/guarded median-latency ratio is 4.50x, and hidden execution exposes 74.3% less parent-visible context than visible execution. Patch validation shows no demonstrated speedup; zero-delay runs expose overhead. End-to-end qualification thus detects lower-layer misses while preserving localization's parallelism.

cs.AI · cs.SE

版本更新 · v6arXiv:2604.23716

Non-Stationarity Breaks Permutation Surrogates in Multi-Agent Reinforcement Learning: Diagnosis and Remedies

Nikolaos Al. Papadopoulos, Konstantinos E. Psannis

原文摘要

Reporting guidance for information-theoretic measures is rarely tested against ground truth. We test one guardrail in two multi-agent reinforcement learning games, a social dilemma and a coordination race, over 100 seeds, on agent pairs whose directed influence is zero by construction. Omitting one precondition, the exclusion of the non-stationary training transient, gives false-positive rates of 100.00% and 99.95%, flagging agents that anneal exploration independently, in runs that never met, as influencing one another. Excluding the transient reaches 3.05% in the social dilemma but 11.8% in the coordination game, so exclusion is neither the only treatment nor a sufficient one. The remedy we recommend changes only the null model, permuting the source within blocks of training time, and reaches 5.25% and 5.50%. Of the three remedies compared (exclusion, this surrogate and conditioning on training time), only the surrogate stays at or below the size of the test in both games. Restricted permutation is an established construction, and the contribution lies in measuring it against ground truth. A third null control, four learners in the social dilemma, defeats all three, and raises the rate of conditioning by 15.20 points against 4.55 for ours. With injected links of known strength, the surrogate detects 82.5% where conditioning detects 58.0% on identical pairs, while the test without exclusion reports a positive with or without a link, so its apparent sensitivity is uninformative. Five block counts from 16 to 256 land in the nominal region, and code and data are released.

cs.AI · cs.IT · cs.LG · cs.MA

新论文 · v1arXiv:2611.23608

JEPA as a Transferable Optimizer: Learning and Understanding Black-Box Optimization from Reward-Agnostic Pretraining

Mayukh Debnath, Manan Roy Choudhury

原文摘要

Black-box optimization evaluates candidates through simulations, rollouts, or experiments, often without gradients. Learned optimizers are usually judged by whether they outperform existing methods. We ask a complementary question: what information does a learned optimizer use from its search history? We introduce JEPA-PS (JEPA Policy Search), based on a Joint-Embedding Predictive Architecture. It is pretrained on random rollouts using parameter vectors and scalar returns. Its latent-prediction objective predicts the next point in an optimization trajectory rather than its reward or gradient, allowing us to examine what was tried and how well it performed. At test time, JEPA-PS reads a short evaluation history and proposes a candidate in one forward pass, without task descriptions, reward shaping, or test-time gradients. We evaluate it on 15 control environments against 12 optimizers, including evolutionary methods, Bayesian optimization, CMA-ES, and an LLM optimizer. JEPA-PS is competitive with tuned classical baselines and outperforms the LLM optimizer by more than 2x on Swimmer at 71x lower per-step cost. It also transfers zero-shot to environments excluded from pretraining. Null-model comparisons and decoder-training ablations indicate that its latent representation carries information beyond simple history statistics; every trained decoder variant performs worse than the default frozen representation. Our main finding is that the history JEPA-PS uses depends on the task. On smooth landscapes it uses the full trajectory, whereas on rugged landscapes the best-seen point carries nearly all the signal. This memory requirement has a weak positive correlation with fitness-distance correlation (FDC; r=0.112, n=7), while an independent reward-ordering sensitivity probe does not corroborate the relationship (r=-0.024, n=7). Code: https://github.com/mdebnath22/jepa-props.

cs.LG · cs.AI

版本更新 · v2arXiv:2611.11826

How Path Integral Diffusion Decides a Finite Spin Glass

Michael Chertkov

原文摘要

When an AI diffusion-type generator samples a frustrated spin glass, which macroscopic alternatives does it decide first, and when? We study this for bridge-diffusion models of the Path Integral Diffusion (PID) type started from a point source. We show that every isotropic PID is a time-changed Gaussian stochastic localization, and conversely every signal-to-noise schedule is realized by a PID protocol. What a continuous-time generator decides therefore does not depend on the protocol, which only sets when and at what cost. We describe the generator's state by its partial commitments: the probabilities it assigns, at each moment, to nested classes of configurations fixed by the target. The partial commitments are martingales that move only along class contrasts. Their ensemble average gives the information about each level at an exact rate set by squared magnetization contrasts. In the balanced, effectively binary approximation, a split reaches half of its class information near γ=1.044/λ_P, where λ_P is the leading eigenvalue of the parent covariance. The theory is applied to Bethe-lattice and Sherrington--Kirkpatrick (SK) glasses with up to 24 spins. It is observed generically that a trajectory first shifts its weight to one of the two mirror halves of the zero-field measure and then to the glassy classes inside it. These classes appear near the glass transition, keep their identity on cooling and are absent in a ferromagnet on the same graphs. Sparse (Bethe) and dense (SK) glasses differ in how often and how strongly such classes exist, not in how trajectories commit to them. For a generator that takes finitely many steps, the protocol matters through the late, local stage rather than the macroscopic decisions. Sampled posteriors at N=32,64 and small trained denoisers show the same sequence.

cond-mat.dis-nn · cond-mat.stat-mech · cs.AI · stat.ML

版本更新 · v2arXiv:2604.17283

HorizonBench: Long-Horizon Personalization with Evolving Preferences

Shuyue Stella Li, Bhargavi Paranjape, Kerem Oktar, Zhongyao Ma, Gelin Zhou, Lin Guan, Na Zhang, Sem Park 等 12 位作者

原文摘要

Personal assistants now talk with the same user for months, and users' preferences change as their lives change; an assistant that keeps serving a preference the user has outgrown fails that user. Real conversations do not record when or why a preference changed, and existing benchmarks either hold preferences fixed or have the user state the change, so they cannot tell whether a model forgot what the user said or failed to update its belief about the user. We introduce HorizonBench and its data generator, which samples how a user's preferences change and generates six months of conversation from that sequence; each change follows a life event that the user mentions without stating the new preference. A human longitudinal preference dataset that we collect and release, 83,024 ratings from 160 people over 13 weeks, sets how often and how much life events change preferences. HorizonBench contains 4,245 questions from 360 users, and when a preference has changed, its earlier value is one of the answer options. All 25 frontier models we evaluate choose the outdated preference more often than evenly spread errors would predict, and the best reaches 52.8% accuracy. Two of the newest models score no higher when given only the relevant conversations than with the full history, yet one added sentence stating the change raises their accuracy by 11 to 22 points, and retrieval and memory methods do not recover the lost accuracy. Remembering what a user said and knowing what the user prefers now are therefore separate capabilities, and HorizonBench, its generator, and the human dataset support research on both.

cs.CL · cs.AI

新论文 · v1arXiv:2611.23601

FutureBench: A Renewable Framework for Evaluating Knowledge-Context Conflict in LLMs

Yi-Cheng Lai, Hen-Hsen Huang

原文摘要

Asking a model to check retrieved text can also tell it which source to trust. We introduce FutureBench, a renewable framework for evaluating knowledge-context conflict using timestamped questions, controlled passage triplets, and no-context probes. Across 3,629 questions and six models, Verify (a prompt combining checking with passage priority) raises false-answer adoption in all five models tested on pre-cutoff events, mainly by displacing answers that were correct under Baseline, which allows both passage evidence and model knowledge. On post-cutoff events, only o4-mini increases adoption (+13.7 percentage points). A four-prompt comparison on 902 post-cutoff questions finds that passage priority without checking raises adoption, while moving to Verify lowers it in all five models. Adversarial-Aware, which warns of misleading passages and prioritizes model knowledge, restores 82.5-98.9% of displaced pre-cutoff answers with under one point of change in true-passage accuracy. FutureBench supports repeated evaluation on fixed cohorts and extension to new events as models change.

cs.CL · cs.AI

新论文 · v1arXiv:2611.23600

Entropy-Regularized Integrated Gradients for Stable Feature Attribution

Rukmangadh Sai Myana, Sumit Kumar Jha, Yanzhao Wu

原文摘要

Path-based attribution methods like Integrated Gradients (IG) are widely used to interpret neural networks, yet imperceptible input perturbations can drastically alter their attribution maps while leaving model predictions unchanged. We introduce Entropy-Regularized Integrated Gradients (ERIG), a path-based attribution method that intrinsically attenuates gradient irregularities along the integration path using maximum-entropy softmax weighting with temperature parameter $τ$. We quantify the effect of the entropy-regularized weighting on attribution stability by using the Lipschitz properties of the model and its gradients. We decompose the attribution change under an input perturbation into a model-response term, whose bound is $τ$-independent, and a weight-variation term, whose bound decreases at rate $\mathcal{O}(1/(2τ))$. Their ratio $R(τ)$ characterizes the trade-off between gradient information and attribution stability. We further add a new dimension to attribution evaluation by adapting classical noise estimators into two metrics, LINV-C and LINV-P, that quantify attribution noise. Across eleven CNN and Transformer backbones on ImageNet and Oxford-IIIT Pet datasets, our experiments show that, relative to the next-best method, ERIG remains comparable on ranking-based fidelity metrics while improving localization by approximately 24% and attribution stability by approximately 47-fold on average, reaching up to four orders of magnitude improvement over IG, all at negligible additional computational cost.

cs.LG · cs.AI · cs.CV

新论文 · v1arXiv:2611.23588

ARISE: Practical Self-Evolving Memory for Accurate and Cost-Efficient Enterprise Agents

Jiangyun Zhang, Zhou Tang, Kristen Surrao, Zhiyu Chen, Albert Guan, Tor Nitayanont, Yupei Zhang, Roopali Singh 等 15 位作者

原文摘要

Enterprise LLM agents operate in dynamic environments where foundation models evolve rapidly, workflows update frequently, and deployments need to satisfy accuracy and cost requirements. We propose ARISE (Agent Refinement through Iterative Self-Evaluation), a practical memory architecture in which the agent's memory evolves from its own failures without updating model weights, tools, or workflow logic. ARISE uses the model itself or a teacher model to attribute the failures and propose corrections. Then, ARISE keeps the corrections as retrievable micro memory, consolidates them into macro rules, and accepts a rule update only if held-out accuracy rises. Across 14 enterprise-workflow domains from SOP-Bench and $τ^2$-bench using 5 models, ARISE-Teacher, which uses Claude Opus 4.6 for attribution and consolidation, gains 10.9 pp over the baseline on SOP-Bench and 5.8 pp on $τ^2$-bench. ARISE-Teacher also outperforms GEPA by 4.5 pp on $τ^2$-bench and 1.4 pp on SOP-Bench. With the deployment model as its own attributor, ARISE-Self still gains 10.0 pp on SOP-Bench. ARISE-Teacher uses 4.0% fewer model calls and 5.2% fewer output tokens at comparable per-task cost (-1.0%).

cs.AI · cs.MA

版本更新 · v4arXiv:2601.16823

Disentangling generalization and memorization in large language models using chess

Leonard S. Pleiss, Maximilian Schiffer, Robert K. von Weizsaecker

原文摘要

Large Language Models (LLMs) exhibit remarkable capabilities, yet it remains unclear to what extent these reflect sophisticated recall or genuine reasoning ability. We introduce chess as a controlled testbed aimed at disentangling these faculties. Leveraging the game's structure and scalable engine evaluations, we construct a taxonomy of positions varying in density of relevant priors -- ranging from common states solvable by memorization to completely novel ones requiring generalization. Crucially, our approach achieves this distinction without requiring explicit knowledge of the models' training data. Applying this taxonomy, we combine a longitudinal analysis of the GPT lineage with a rigorous evaluation of contemporary models, including Claude Opus and Gemini. Our analysis reveals a steep gradient: performance consistently degrades as the density of relevant priors decreases. Notably, for tasks with few relevant priors, base model performance regresses to the random-play baseline, and all models propose illegal moves at sharply increasing rates, indicating that they fail to reliably apply the fixed rules of the game in unfamiliar positions. While newer models improve, progress slows significantly for tasks with sparse priors. Furthermore, while reasoning-augmented inference improves performance, its relative marginal benefit per token decreases in the absence of relevant priors. These results suggest limitations in systematic generalization, highlighting the need for mechanisms beyond scale to achieve robust performance when deprived of relevant priors.

cs.CL · cs.AI

新论文 · v1arXiv:2611.23581

ExiST: Existence-Aware Self-Training for Semi-Supervised Multi-Organ Segmentation

Maxx Richard Rahman, Prakhar Kumar, Wolfgang Maass

原文摘要

Semi-supervised multi-organ segmentation should learn more than a dozen organs from a handful of labeled scans, and the smallest organs occupy less than one percent of the foreground. We identify a failure mode of self-training that is specific to this regime. When a teacher misses a small organ in a scan, every voxel of that organ receives the confident pseudo-label background, and the student is trained toward the organ absence. Repeated over many scans, this loop extinguishes tail organs. In a pilot study, removing the gallbladder from the pseudo-labels reduced its Dice on AMOS from 36.5 to 1.0. We argue that a missing organ should be treated as unresolved rather than absent and propose Existence-aware Self-Training (ExiST). It builds pseudo-labels by leave-one-out consensus of independently trained networks, screens them with a quality gate on validation scans, and infers, per scan and per organ, whether the organ was found. For every unresolved organ, the background voxels of that scan receive the set-valued target, trained with a marginal cross-entropy that accepts either answer and a Dice term that skips uncertain voxels. We show that this loss never pushes a candidate organ toward absence and never favors one member of the set, which leaves the decision to labeled evidence. ExiST outperforms the current state-of-the-art methods and reaches 77.1 mean Dice on Synapse with 20% labeled scans and 75.9 on AMOS with 5% labeled scans.

cs.CV · cs.AI

新论文 · v1arXiv:2611.23572

Receiver-Conditioned Clarification for Multi-Agent Food Assistance

Kaiwen Shi, Zheyuan Zhang, Changzhi Liu, Han Bao, Yanfang Ye

原文摘要

Food insecurity creates a pressing need for guidance that helps households find usable food resources and assess food choices in light of their circumstances. Such guidance must connect eligibility requirements, practical access constraints, and dietary needs rather than return isolated listings or generic nutritional advice. Multi-agent food assistants can coordinate these tasks, with uncertainty indicating when clarification or verification is needed before a recommendation. However, uncertainty can be lost across agent handoffs, allowing tentative interpretations to propagate as facts and obscuring opportunities for correction. Moreover, confidence that a service requirement has been interpreted correctly is not confidence that a recommendation satisfies the user's needs; local confidence cannot simply be reused as downstream reliability. We introduce an inference-time, receiver-conditioned clarification framework that preserves shared premise identities and replays consistent interpretations across dependent agents. Each receiver exposes admissible decisions and remaining evidence obligations, so question relevance reflects both decision changes and missing support. Information questions and user-confirmable fallback adjustments are evaluated together with their anticipated follow-up interactions under shared budgets. This structure connects clarification to jointly supported plans for resource navigation, nutritional reasoning, and supply-side planning.

cs.AI

新论文 · v1arXiv:2611.23567

Panini- Feasible Grammar For Robot Designs

Mayukh Debnath, Manan Roy Choudhury

原文摘要

Symbolic methods make robot designs interpretable but often overlook continuous geometry; learned methods handle geometry but offer less transparent structure. We introduce a typed production grammar that uses case-role relations, parameter inheritance, and specificity-based rule arbitration to generate both manipulation skills and robot morphologies. A geometric realization map translates grammar composition into transformations in SE(3), behaving as a monoid homomorphism when inheritance is inactive and exhibiting a measurable defect when inheritance is enabled. The grammar provides a compact representation, and its arbitration decisions can be recovered from geometric containment (agreement 0.905 +/- 0.062 across 10 seeds, versus 0.611 +/- 0.090 for a Euclidean control). We also introduce PANINI, a cross-entropy search method over grammar-based morphology representations. Across 30 methods and 24 seeds, PANINI achieves the best mean rank but is not statistically distinguishable from strong generic optimizers. Its main demonstrated advantage is therefore structural rather than optimization performance: grammar-generated designs are legal by construction, and coordinated limb placement produces 95% jointly feasible edits, compared with 55% for range-constrained resampling and none for flat parameterization in 6,000 draws. These results show how typed grammars can make robot generation more interpretable and geometrically coordinated, without implying that every feasible design will locomote or that grammar-based search universally outperforms generic optimizers.

cs.RO · cs.AI

版本更新 · v3arXiv:2602.15198

Colosseum: Auditing Collusion in Cooperative Multi-Agent Systems

Mason Nakamura, Abhinav Kumar, Saswat Das, Sahar Abdelnabi, Saaduddin Mahmud, Ferdinando Fioretto, Shlomo Zilberstein, Eugene Bagdasarian

原文摘要

Multi-agent systems, where LLM agents communicate through free-form language, enable sophisticated coordination for solving complex cooperative tasks. This surfaces a unique safety problem when a group of agents forms a coalition and colludes to pursue secondary goals and degrade the joint objective. In this paper, we present Colosseum, a framework for auditing LLM agents' collusive behavior in multi-agent settings. We ground how agents cooperate through a formal multi-agent decision-making framework and measure action-based collusive behavior in actions via regret relative to the cooperative optimum and compare it with communication-based collusive behavior. Colosseum enables audits of LLM agents for collusion under benign settings, different coalition objectives, persuasion tactics, and network topologies. We then introduce a new behavioral probe by creating secret communication channels between agents, showing that most out-of-the-box models exhibit a propensity to collude under this probe, which we term emergent collusion. Furthermore, we discover ``collusion on paper'' when agents plan to collude in text but often pick non-collusive actions. Colosseum provides a new way to audit collusion in cooperative multi-agent systems while presenting observations about how collusion emerges, what affects collusion efficacy, and which strategies may mitigate it.

cs.MA · cs.AI · cs.CL

新论文 · v1arXiv:2611.23564

LProp: Propagating Latent Uncertainty through Evidence-Conditioned Belief Revision in LLM Agents

Kaiwen Shi, Zheyuan Zhang, Weixiang Sun, Han Bao, Colby Nelson, Yanfang Ye

原文摘要

When LLM agents pass intermediate judgments downstream without uncertainty, later decisions may treat tentative assumptions as facts. Effective uncertainty propagation must preserve information for correction and update inherited beliefs as new evidence arrives. Yet scalar scores omit information needed for revision, while detailed descriptions and reflection incur additional costs. We introduce Latent Uncertainty, a latent-space estimator that preserves uncertainty-relevant information without lengthy descriptions. Building on this, we propose LProp (Latent Uncertainty Propagation), which revises inherited answer distributions using latent representations and new evidence without additional trajectory sampling. We evaluate both methods across multiple-choice, yes/no, open-ended question answering, and mathematical reasoning tasks. Our approach outperforms baselines on most metrics, while LProp achieves the lowest mean inference-token usage in all eight propagation settings. Matched feature ablations further demonstrate that inherited latent content improves error recovery beyond answer probabilities or model capacity alone.

cs.AI

新论文 · v1arXiv:2611.23560

FACET - Facet Aware Contrastive Embeddings for Traits- Big Five Estimation from Life Narratives

Balakrishnan Arumugam, Priyen Parekh, Mehak Gupta, Joshua Oltmanns

原文摘要

Personality is only weakly and inconsistently signaled by written language, and existing text-based methods typically combine language-model and psycholinguistic representations at the input or decision level. It remains open whether a more deliberate combination of the two views, and a decomposition of the prediction target, can recover more of the signal. We present a three-phase framework for estimating Big Five personality domain scores from life-narrative interview text. It encodes text and psycholinguistic features separately and combines them by concatenation or by a learned weighting, decomposes each domain into contrastively trained facet embeddings instead of predicting the domain directly, and pools the facet embeddings by attention into a domain-level prediction. We compare end-to-end with staged training and ablate the input modalities, their combination, and the facet-pooling strategy. On 1,332 participants from the St. Louis Personality and Aging Network (SPAN) study, facet embeddings reach AUC up to 0.740, and domain-level R2 reaches 0.328, highest for Openness. End-to-end training gives lower error than staged training in every domain, and attention pooling gives the lowest error in every domain. Combining the two views lowers error relative to text alone for Conscientiousness and Agreeableness, whereas a learned weighting of the views shows no consistent advantage over concatenation. The framework provides facet- level predictions that a domain-only model cannot, and it improves modestly on a regression from the raw features in four of the five domains.

cs.AI

版本更新 · v3arXiv:2603.13669

SHAMISA: SHAped Modeling of Implicit Structural Associations for Self-supervised No-Reference Image Quality Assessment

Mahdi Naseri, Zhou Wang

原文摘要

No-Reference Image Quality Assessment (NR-IQA) aims to estimate perceptual quality without access to a pristine reference. Learning an NR-IQA model faces a fundamental bottleneck because it requires many costly human perceptual labels. We propose SHAMISA, a non-contrastive self-supervised framework that learns from unlabeled distorted images through explicitly structured relational supervision. Unlike prior methods that impose rigid binary similarity constraints, SHAMISA shapes soft, controllable implicit structural associations that are both distortion-aware and content-sensitive. A key innovation is its compositional distortion engine, which generates an uncountable family of degradations from continuous parameter spaces and groups the resulting images so that exactly one distortion factor varies at a time. This construction enables fine-grained control over representational similarity during training. Images with shared distortion patterns are pulled together in the embedding space, while severity variations produce structured shifts. Dual-source relation graphs integrate known degradation profiles with emergent structural affinities throughout pre-training. The learned encoder is then frozen and paired with a linear regressor. Extensive experiments on synthetic, authentic, and cross-dataset NR-IQA benchmarks demonstrate strong overall performance, improved transfer in the reported cross-dataset comparisons, and favorable qualitative robustness diagnostics. Self-supervised pre-training requires no IQA opinion-score annotations or contrastive loss.

cs.CV · cs.AI · cs.LG

新论文 · v1arXiv:2611.23546

Language Models Make New Research Sound Old

Guy Zyskind

原文摘要

Researchers are increasingly using large language models (LLMs) to validate research questions. We find that LLMs make open research questions sound partly studied and overstate what related papers establish. Models recognize most of these overstatements when asked to check the papers' full texts separately. On 40 questions drawn from newly accepted papers, GPT-6-Astra correctly judges 98% to be unstudied, but states that status plainly in only 8% of ordinary answers. All six frontier systems show reporting gaps of 30-90 percentage points. Across 22 models given identical papers, higher capability is associated with more accurate judgments and larger reporting gaps. This reporting failure is asymmetric: models correctly identify studied questions in 88-100% of ordinary answers when we supply the answering papers. To correct this, we build a fixed answer from each model's judgment and separate checks of cited papers. This raises Astra's clear reporting of open status from 8% to 95%, with improvements in all six systems. Finally, we predict reviewers' novelty grades by combining several models' written ratings with a rule learned from graded examples, reaching the performance range reported by Think-Probe-Respond. Unlike TPR, our approach needs no access to model weights or internal activations, so it also works with closed-weight frontier models.

cs.CL · cs.AI

版本更新 · v2arXiv:2610.34747

Query-Driven Learning for Lifted Reasoning

Luise Ge, Brendan Juba

原文摘要

Learning and reasoning are both essential to intelligent agents, but have traditionally been treated as two separate stages: first learning a knowledge base from data (or taking one as given), and then verifying queries as if the knowledge base were complete. This separation has been shown to be inefficient when the downstream goal is to verify only a small set of queries. Moreover, it neglects the possibility that data and a knowledge base contain partial yet complementary information, requiring an agent to combine both sources to justify or refute a query. Motivated by the need to integrate learning and reasoning more tightly, we study the problem of query-driven learning for reasoning. More specifically, we extend prior work to the first-order setting, which is critical for practical applications as the number of ground atoms typically grows rapidly. We begin by asking when validating or refuting a universal query given partial data and knowledge base is even possible, and identify a necessary and sufficient condition called bounded concealment. We then propose a quasi-polynomial-time algorithm that solves the task at the lifted level. Finally, we establish soundness and completeness guarantees relative to space-bounded lifted proofs, along with quasipolynomial sample and runtime bounds, demonstrating that lifted query-driven reasoning can be efficient even when unrestricted knowledge-base learning is not.

cs.AI

版本更新 · v2arXiv:2605.07058

MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments

Yicheng Gao, Xiaolin Zhou, Yahan Li, Yue Zhao, Ruishan Liu

原文摘要

Real-world clinical diagnosis is a complex process in which the doctor is required to obtain information from both interaction with the patient and conducting medical exams. Additionally, the doctor needs to adapt to different patient personas, as well as noisy and incomplete information that can happen at any time during the process. However, existing benchmarks for medical LLMs and methods for automatic diagnosis largely simplify this process by reducing it to single-turn question answering, noise-free conversations, or sequential exam making, etc., ignoring the interactive and uncertain nature of clinical diagnosis. In this paper, we aim to address this gap by formalizing clinical diagnosis as a Partially Observable Markov Decision Process (POMDP) with three action types: questioning the patient, ordering medical exams as tool calls, and issuing a diagnosis. We also introduce a systematic noise model comprising seven patient noise types and three exam noise types. Using our proposed environment, we train an effective diagnosis agent, \textbf{MedExAgent}, through a two-stage pipeline that first performs supervised finetuning on synthetic conversations structured after the Calgary-Cambridge model for clinical interviews, and then applies DAPO to optimize a composite reward capturing diagnostic accuracy, tool call quality, and exam cost including financial cost and patient discomfort. Through extensive experiments and ablation studies, we demonstrate that MedExAgent achieves diagnostic performance comparable to larger models while maintaining cost-efficient examination strategies.

cs.CL · cs.AI

新论文 · v1arXiv:2611.23520

BeFit: Shifting Complexity from Weights to Bernstein Activations for Resource-Constrained Sensing

Malak Gamal El-Din, Yasser Shoukry, Salma Elmalaki

原文摘要

Time-series sensing increasingly relies on neural networks, yet deployment on resource-constrained microcontrollers remains challenging. Aggressive model compression can reduce accuracy, while restricting weights to multiplication-free representations introduces further losses. We present BeFit, an end-to-end framework that shifts model capacity from weights to learnable Bernstein polynomial activations. BeFit (i) finds compact Bernstein-activated convolutional neural networks (BeCNNs) through hardware-aware neural architecture search and trains them via knowledge distillation; (ii) projects their weights to multiplication-free codes, freezes the codes, and recovers the resulting accuracy loss through the learnable activations; and (iii) generates model-specific kernels that exploit the projected weights for multiplication-free convolutions. On human activity recognition, BeCNNs are 2.75 percentage points (pp) more accurate on average than ReLU CNNs of identical architecture. On a TI CC3220SF (Cortex-M4), weight projection speeds BeCNNs up by 2.2x on average while losing 0.8 pp, compared with 2.4 pp for ReLU CNNs. Against the MLPerf Tiny DS-CNN reference model running on TensorFlow Lite Micro with CMSIS-NN, the unprojected BeCNN achieves 1.0 pp higher accuracy, 2.2x lower latency, and 6.4x lower flash, while the projected model remains within 1.0 pp of the reference accuracy and runs 5.9x faster. In emulation on a RISC-V core without a hardware multiplier, the instruction-count reduction relative to the DS-CNN increases from 5.2x on the Cortex-M4 to 147x. BeFit also extends to transformers and MLPs, and across all tasks the final models stay within 1.56 pp of their float32 teachers on average.

cs.AI · cs.AR

新论文 · v1arXiv:2611.23516

Beyond the Sink Token: Sink-Guided Key Editing in Vision-Language Models

Hui Zhu, Saleh Zare Zade, Rafi Ibn Sultan, Xiangyu Zhou, Dongxiao Zhu

原文摘要

In vision-language models (VLMs), a few sink tokens attract a disproportionate share of attention. While previous studies have examined why these tokens attract attention, we investigate how the sink-key direction affects attention to image and question tokens. Across four probed VLMs, boosting the magnitude of positive sink-aligned components increases attention on average, whereas boosting negative components decreases it; these responses persist through later layers. This identifies the native sink key as a reference for controlling attention beyond its own position. Motivated by this finding, we propose SHRINK, a training-free method that partially reduces sink-aligned components across visible keys during generation. On nine benchmarks under sampled reasoning, it improves mean accuracy by 5.1 and 2.7 percentage points on LLaVA-OneVision-7B and Qwen2.5-VL-3B, with smaller gains on the other two main models, Qwen2.5-VL-7B and Gemma-3-4B. Beyond SHRINK, rescaling sink-aligned key components can replace the attention operations of two existing attention-steering methods, GIFT and ACG, with comparable accuracy on hallucination benchmarks.

cs.CV · cs.AI