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

+549新增论文
+443其中 AI 相关
+62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 18:43 — 2026-11-21 18:43(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:2611.21682

NEBench R1: a Reasoning Benchmark for Evaluating Nuclear Domain Competence in Large Language Models

Liyan Wang, Longze Li, Mohamed Hassan, Isshu Lee, Dajie Sun, Ruimin Feng, Aleksandar Vakanski

原文摘要

Large language models (LLMs) are increasingly being explored for knowledge-intensive and safety-critical applications in nuclear engineering. However, existing nuclear-domain benchmarks provide limited content coverage or focus primarily on answer accuracy without systematically evaluating the quality of the underlying reasoning. This study introduces NEBench R1, a reasoning-oriented benchmark comprising 7,546 unique multiple-choice questions extracted from U.S. Nuclear Regulatory Commission (NRC) reactor operator licensing examinations covering 18 nuclear facilities. Each question is classified according to the NRC Knowledge and Abilities catalogs and accompanied by an official explanation serving as a reference for reasoning evaluation. Six LLMs were evaluated, with overall accuracy ranging from 34.0% to 73.9% across three distinct performance tiers. To assess reasoning quality, an LLM-based judge evaluated correctly answered responses using a multidimensional rubric. Although responses generally demonstrated strong problem comprehension, they scored lower in factual and procedural accuracy and distractor analysis, indicating that correct answers were not necessarily supported by technically sound justifications. Response clarity was also weakly indicative of technical quality. Furthermore, longer reasoning traces were associated with lower accuracy in both frontier reasoning models, suggesting that reasoning length may reflect question difficulty rather than improved problem-solving performance. Reasoning-token consumption may therefore provide an additional signal for identifying potentially unreliable responses. NEBench R1 establishes a benchmark for evaluating both answer accuracy and reasoning quality, providing a foundation for assessing the domain competence and reliability of LLMs in nuclear engineering.

cs.AI

版本更新 · v2arXiv:2604.15559

Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation

Jacob Dang, Brian Y. Xie, Omar G. Younis

原文摘要

Recent work on subliminal learning has shown that language models can transmit semantic traits through data with no apparent connection to those traits. Whether this phenomenon extends to agentic systems, where policies are learned from trajectories rather than static text, remains an open question with critical safety implications. We present the first empirical evidence that unsafe agent behaviors transfer subliminally through model distillation, demonstrated across several settings. In the primary setting, we construct a teacher agent with a deletion bias, a tendency to perform destructive file-system actions through an API-style tool interface, and distill it into a student using only trajectories from ostensibly safe tasks, with all explicit deletion keywords filtered. In the secondary setting, we replicate the threat model with shell commands, replacing API calls with shell commands and operationalizing the bias as a preference for chmod over semantically equivalent alternatives (e.g., chown, setfacl) when issuing the first permission-related command. Despite sanitation, students inherit measurable biases. Over 1,500 trials per configuration, the API student chooses a destructive action first on 35.5% of ambiguous tasks (vs. 1.2% baseline), and the Bash student's chmod-first rate reaches 30-55% (vs. 0-10% baseline). In a multi-turn environment where tool calls execute, 33.9% of student episodes end with a resource irreversibly destroyed (vs. 6.4% baseline). On TerminalBench we observe a smaller effect (7.7% vs. 0% baseline). Explicit data sanitation is therefore insufficient as a defense; mitigating behavioral bias transfer in agentic systems will require fundamentally new strategies.

cs.AI

新论文 · v1arXiv:2611.20195

How Specification Explicitness Shapes Failure in LLM-Generated Code

Elmira Onagh, Maleknaz Nayebi

原文摘要

When LLMs turn requirements into code, every underspecified constraint becomes behavior that the model must decide during generation. We test when specifications move beyond structure and begin to control generated behavior. In a controlled experiment, 72 participants implemented and revised Python tasks with an LLM under three levels of specification explicitness. The results show a clear break between lightweight structure and explicit behavioral obligations. Lightweight structure reduced hidden-test failures by only 8.7%. Requiring participants to specify constraints, invariants, negative requirements, edge cases, tie-breaking rules, and exception handling cut failures by 47.2% and reduced the number of distinct failure modes by 20.5%. The largest drops were in failure modes tied to implicit obligations. Invariant violations fell by 66.1%, and unsupported assumptions fell by 59.5%. After a requirement change, specifications with explicit behavioral obligations produced regressions in more submissions, but each affected submission lost less previously correct behavior. Explicit specifications also changed how developers judged generated code. Without explicit obligations, participants overestimated generated-code correctness by about 23 percentage points. With explicit obligations, they produced better code but judged it more cautiously. Thus, the lesson is sharper than "better specifications help." Lightweight structure barely helped, while explicit behavioral obligations nearly halved failures, made regressions shallower, and improved calibration.

cs.SE · cs.AI

版本更新 · v4arXiv:2604.24197

Seeing Is No Longer Believing: Frontier Image Generation Models, Synthetic Visual Evidence, and Real-World Risk

Shuai Wu, Xue Li, Zhijun Wang, Bolun Liu, Weilin Cai, Zihao Su, Ran Wang

原文摘要

Image generation systems can produce plausible photographs, readable documents, and consistent depictions of people and places. Presented as records of real events, these outputs can influence decisions in news, finance, identity verification, medicine, and law. This narrative review examines public model documentation, incident reports, empirical studies, governance materials, and illustrative English and Chinese practitioner commentary available through 1 October 2026. We separate vendor capability claims, documented incidents, experimental findings, and prospective harm pathways. Realism, text rendering, reference consistency, conversational editing, search grounding, and low generation cost shape the conditions under which synthetic images acquire evidentiary weight. Documented episodes show how fabricated visuals enter public and institutional workflows, although public reporting cannot measure misuse prevalence for current releases. Comparing provider safeguards, provenance credentials, watermarks, platform labels, and statutory duties shows that origin signals do not verify a depicted event or transaction. Our framework matches artifact types and decision stakes to specific controls. Consequential decisions require authenticated source records, independent channel corroboration, and proportionate review before action.

cs.CL · cs.AI

新论文 · v1arXiv:2611.20187

Cyber-Secure Wearable, Ultra-Low-Power Networked Sensors for Soldier Health Monitoring: Insights from the CAPSARII project

Luciano Bozzi, Christian Celidonio, Umberto Nuzzi, Massimo Biagini, Stefano Cherubin, Asbjørn Djupdal, Tor Andre Haugdahl, Andrea Aliverti 等 22 位作者

原文摘要

The CAPSARII project advances a cyber-secure, ultra-low-power wearable platform and Internet of Battlefield Things (IoBT) framework to support real-time monitoring of military personnel and data-driven decision support. The architecture integrates smart-textile sensing, a body/squad area mesh, edge-deployed AI on secure RISC-V microcontrollers, and cloud/federated analytics for cross-mission learning and comparative studies. Two representative use cases--pilot fatigue and warfighters operating in harsh environments--guide requirements, validation, and interface design (including HUD and haptic feedback). As an initial step toward embedded models, we benchmark stress classification on a public wearable dataset: tree-ensemble baselines achieve state-of-the-art performance for accuracy and F1 (e.g., XGBoost: $\mathrm{Accuracy}=0.901\!\pm\!0.023$, $\mathrm{F1}=0.920\!\pm\!0.019$), while a feature-based Transformer attains the highest AUC ($0.963\!\pm\!0.025$), informing feature-engineering and compression choices for edge deployment. Building on these results, CAPSARII employs HW/SW co-design, model compression (pruning, quantization, distillation), and precision tuning to meet stringent energy and memory budgets; it further investigates Processing-in-Memory accelerators for future efficiency gains. A structured gap analysis highlights the remaining challenges: establishing validation standards for physiological signal quality and textile durability; achieving low-latency, scalable BAN/SAN communications under mobility and interference; and deploying hybrid lightweight/post-quantum cryptography with robust device identity and side-channel resilience on constrained endpoints. By consolidating these elements into a coherent, secure IoBT stack, CAPSARII delineates a practical roadmap from laboratory baselines to mission-ready soldier monitoring in both training and operational scenarios.

cs.ET · cs.AI · cs.AR · cs.CR · cs.HC · cs.LG

新论文 · v1arXiv:2611.20185

RoboKit: Accelerating Parallel Differentiable Optimization for Robot Learning

Jianglong Ye, Xuanbin Peng, Conor Mc Gartoll, Changwei Jing, Lai Wei, Jiali Chen, Xueyan Zou, Xiaolong Wang

原文摘要

We present RoboKit, a high-performance, general-purpose toolkit for parallel differentiable optimization in robotics, built on three principles: (i) Performance and flexibility: GPU-parallel multi-seed solvers pair sampling with versatile optimizers, running 2-60$\times$ faster than task-specific baselines across seven tasks; (ii) Geometry, rendering and physics: hybrid and coarse-to-fine representations accelerate collision, differentiable rendering and simulation, optimizing 4096 grasps in 2.97 s; (iii) Integration with learning: costs are differentiable and simulation is parallel, so RoboKit both distills optimization into policies and trains RL; distilled policies run 9-240$\times$ faster at comparable accuracy, and RL reaches a GPU baseline's scores up to 2.1$\times$ sooner on the same simulator.

cs.RO · cs.AI · cs.LG

版本更新 · v3arXiv:2607.11250

Do Multi-Agent LLMs Explore Each Other?

Hyeong Kyu Choi, Jiatong Li, Wendi Li, Xin Eric Wang, Sharon Li

原文摘要

Exploration is essential for reliable autonomy in multi-agent systems, yet it remains unclear whether large language model (LLM) agents can explore effectively when interacting with one another. We show that modern LLM agents often fail to do so, exhibiting myopic and polarized interaction patterns that lead to suboptimal coordination and increased regret. We formalize this challenge as the Multi-Agent Exploration problem, modeling it as a partially observable stochastic game (POSG), in which agents must probe peers to infer their capabilities and identify effective interaction strategies. To address this, we introduce Multi-Agent Contextual Exploration (MACE), a lightweight framework that explicitly promotes exploration through structured peer selection. Across both contextual and parametric diversity settings, exploration behavior and downstream task performance substantially improve when MACE is applied. We further provide theoretical insight that MACE achieves a sublinear expected-regret bound. Overall, our results highlight a fundamental limitation of current LLM agents and underscore the importance of explicitly guided exploration for reliable multi-agent autonomy.

cs.MA · cs.AI

版本更新 · v3arXiv:2605.26321

Anchor: Preventing Artifact Drift for Reliable Agent Verification

Abhijay Rana, Maksim Ivanov

原文摘要

Reliable evaluation of long-horizon agents requires verifiers that match what the task actually asks. Verifiable rewards have driven progress in domains like math and coding, yet building similarly reliable environments for open-ended enterprise workflows repeatedly struggles to balance realism, verifiability, and scale. This creates a validity problem we call artifact drift: when instructions, environments, oracles, and verifiers are created separately, they frequently fail to agree on what a task requires, resulting in environments that are unsolvable, reward-hackable, or inconsistent. We introduce Anchor, a benchmark-construction pipeline that formalizes domain expert specifications of business workflows into constrained optimization programs. From a single parametric specification, the pipeline jointly produces a natural-language instruction, environment configuration, solver-certified ground-truth solution, and state-based verifier. With this pipeline, altering parameters yields new tasks with controlled difficulty and known optimal solutions, producing harness-agnostic environments whose rewards depend solely on end-state business correctness. We apply Anchor to produce ERP-Bench: a benchmark of 300 long-horizon tasks spanning procurement and manufacturing workflows in a production-grade ERP system. Our pipeline organizes tasks into diverse families with calibrated difficulty profiles, and allows generated environments to evaluate coding, browser, and computer-use agents using a shared verifier. Across these harnesses, we find that frontier models follow all rules and instructions in 25.9% of trials and reach fully optimal decisions in only 17.4% of trials. Overall, Anchor and ERP-Bench offer a concrete recipe for building auditable evaluation environments whose verifiers remain aligned with the economically valuable tasks they are intended to score.

cs.AI

版本更新 · v3arXiv:2611.03100

LUCA: A Synthetic World for Experimental Reasoning

Anna Gogleva, Thomas Charman, Martin Balla, Karl Gemayel

原文摘要

[Note, November 2026: the experimental results in this version are affected by implementation issues in the simulation environment and should not be relied upon pending correction. See the author note on page 1.] An experimental agent can succeed for the wrong reasons. Evaluations, however, often focus on the outcome, not the experiments, evidence, and inferences that produced it. We introduce LUCA, a configurable synthetic world for evaluating this process under fully known ground truth. Agents learn unknown regulatory systems through budgeted experimentation, then predict unseen interventions, diagnose hidden faults, and attempt rescues. This setup allows us to evaluate experiment selection, prediction, diagnosis, and intervention separately. Across seven frontier LLMs and three baselines, these capabilities dissociate sharply. Agents frequently rescue malfunctioning systems while diagnosing the wrong fault; diagnosis varies 7x across models while prediction varies only 1.4x. Models also adopt surprisingly different experimental strategies, and greater exploration does not reliably yield better performance. In one failure mode, agents often perform the experiment needed to distinguish competing causes, but still fail to use the resulting evidence. LUCA turns these hidden failures into measurable behaviours, providing a controlled environment to study where experimental agents succeed and fail. Code: https://github.com/RelationRx/luca.

cs.AI

版本更新 · v2arXiv:2611.16013

Runtime Analysis of Cartesian Genetic Programming on Max: A Proven Exponential Speedup

Duc-Cuong Dang, Roman Kalkreuth, Andre Opris

原文摘要

Genetic Programming (GP) is a search paradigm inspired from natural evolution for the automated discovery of expressions, functions, and computer programs. Cartesian GP (CGP) is a flavor of GP that uses a graph-based model to encode candidate programs, as contrasted to the conventional Tree-based GP (TGP). Since its inception two decades ago, CGP has predominantly been analyzed empirically, thus little is known regarding its theoretical performance guarantees. This paper analyzes CGP for the Max problem, which has been only rigorously studied for TGP with a proven expected runtime superlinear in the optimal program length $n$. We prove that the expected time for the ($1$+$1$) CGP employing common mutation operators to evolve an equivalent optimal of the same output is polylogarithmic in $n$, specifically at most $O(\log^9{n})$. This is remarkable, as it shows an exponential speedup by switching to CGP. Our results shed light on the benefit and the compactness of the graph-based representation for programs, and on how CGP can navigate its search and fitness spaces efficiently. This is a first proven performance guarantee for CGP on an established benchmark problem. Experiments complement our theoretical findings.

cs.LG · cs.AI · cs.NE

版本更新 · v2arXiv:2505.11621

Benign Overfitting of Two-Layer ReLU Networks in Fixed Dimension through Localized Initialization

Junhyung Park, Patrick Bloebaum, Shiva Prasad Kasiviswanathan

原文摘要

Benign overfitting occurs when a model fits noisy training data exactly and still generalizes well. We establish this behavior for finite-width two-layer networks with the standard ReLU activation, trained by gradient flow in fixed dimension. Our data-independent initialization combines Gaussian weights with neurons initially active only on shrinking spherical caps. For uniform spherical inputs and any fixed bounded conditional response law, training and excess risks vanish in probability uniformly over all sufficiently late times. The parameters also converge, with probability tending to one, to an interpolating limit whose excess risk vanishes in probability. We use one schedule depending only on sample size and dimension. Our proof connects the network to a kernel flow whose limit combines ridge regression with localized spikes that fit the training residuals. We give explicit sample-size, width and training-time bounds.

cs.LG · cs.AI · stat.ML

版本更新 · v2arXiv:2604.08541

Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts

Haolei Xu, Haiwen Hong, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen 等 10 位作者

原文摘要

Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.

cs.CV · cs.AI · cs.CL

版本更新 · v2arXiv:2607.26057

Pass the Baton: Trajectory-Relayed On-Policy Distillation

Haolei Xu, Xiaowen Xu, Haiwen Hong, Zixuan Ni, Hongxing Li, Yiwen Qiu, Weiming Lu, Yongliang Shen

原文摘要

On-policy distillation (OPD) grounds token-level supervision in the student's own trajectory, yet suffers from prefix failure: once the student commits to a wrong reasoning direction, all subsequent generation builds on this deviation, producing misdirected continuations that elicit unreliable supervision and waste compute. We identify a teacher-student continuation asymmetry on failed prefixes, where the teacher tends to redirect while the student continues along the original direction, and convert it into a label-free handoff trigger in Relay On-Policy Distillation (Relay-OPD). During training, Relay-OPD constructs relay trajectories by letting the teacher briefly take over at detected trigger points to produce a teacher leg, after which the student resumes and is optimized on the resulting trajectory. A limited relay budget concentrates intervention on critical early positions while limiting departure from the student policy. With a Qwen3-4B-Instruct-2507 teacher and Qwen3-0.6B/1.7B-Non-Thinking students on eight mathematical reasoning benchmarks, Relay-OPD achieves the best or second-best results on every benchmark, outperforming standard OPD by +5.73% and the strongest baseline FastOPD by +1.49% on average for 1.7B, with consistent gains at 0.6B. Training trajectory length is reduced by over 50%.

cs.CL · cs.AI

新论文 · v1arXiv:2611.20144

HEAMS: Holistic Evaluation of Agent Memory Systems

Chirag Goel, Jacob Lavoie, Amirhossein Naghshzan, Jordan Bannister, Marzia Nouri, Mehar Bhatia, Matt Harris, Siva Reddy 等 10 位作者

原文摘要

Memory systems allow AI agents to persist information across interactions, making them more personalized and more capable on long-horizon tasks. This persistence introduces new risks: stored information can be disclosed in inappropriate contexts, resist correction or deletion, be retained beyond what users desire, or be poisoned to influence later agent behaviour. Yet memory systems are typically evaluated on task performance and efficiency alone. We argue that a memory system should instead be understood as the custodian of its user's persistent information, with responsibilities that extend beyond helping an agent complete its tasks, and that it should be evaluated holistically against all such responsibilities. We present Holistic Evaluation of Agent Memory Systems (HEAMS), an open-source framework for evaluating agent memory systems in this way. First, we state the responsibilities of agent memory systems as explicit behavioural desiderata spanning performance, efficiency, security, privacy, and user sovereignty. Second, we assemble a benchmark suite that covers these desiderata and acknowledge the gaps that remain. Third, we develop MilaMem, a prototype memory system, and a governance layer that can wrap any memory system, enabling controlled study of memory system design choices. Fourth, we compare memory systems across the full HEAMS benchmark suite. We find that no memory system leads every performance benchmark, and strong performance scores may coexist with poor privacy, user-sovereignty, and security scores. Governance strongly influences results: recall policies can reduce inappropriate disclosure at a cost in helpfulness, while both recall and retention policies improve poisoning resistance. Performance and efficiency benchmarks alone therefore cannot reveal how well a memory system meets its responsibilities. Code: https://github.com/milatechtransfer/heams

cs.AI · cs.CL · cs.CR

版本更新 · v4arXiv:2504.15564

OpenClassGen: A Large-Scale Corpus of Real-World Python Classes for LLM Research

Musfiqur Rahman, SayedHassan Khatoonabadi, Emad Shihab

原文摘要

Existing class-level code generation datasets are either synthetic (ClassEval: 100 classes) or insufficient in scale for modern training needs (RealClassEval: 400 classes), hindering robust evaluation and empirical analysis. We present OpenClassGen, a large-scale corpus of 324,843 Python classes extracted from 2,970 engineered open-source projects. Each entry pairs a human-written class with its corresponding skeleton, which comprises class and method signatures with associated docstrings, and is enriched with 27 static code metrics covering complexity, coupling, cohesion, and inheritance properties. Unlike prior benchmarks that require repository-level context resolution, OpenClassGen provides self-contained class skeletons that serve as complete generation specifications. We demonstrate the corpus's utility by evaluating three LLMs (GPT-o4-mini, Claude-4-Sonnet, Qwen-3-Coder) on a curated, executable subset of 300 classes, enriched with test suites achieving 58% branch coverage. Results show strong semantic similarity (CodeBERTScore-F3: 0.89) but moderate functional correctness (pass rate: 0.33), with substantial variance across models. This variance, along with diverse class characteristics, confirms that OpenClassGen enables meaningful differentiation of LLM capabilities. The dataset supports diverse use cases, including fine-tuning, retrieval-augmented generation, difficulty modelling, and failure mode analysis. The complete dataset and curation scripts are publicly available at https://zenodo.org/records/18409150.

cs.SE · cs.AI · cs.LG

版本更新 · v3arXiv:2510.26130

Beyond Synthetic Benchmarks: Evaluating LLM Performance on Real-World Class-Level Code Generation

Musfiqur Rahman, SayedHassan Khatoonabadi, Emad Shihab

原文摘要

Large language models (LLMs) have advanced rapidly from well-studied function-level code generation to emerging repository-level approaches, yet this progression reveals a critical gap: class-level code generation, which is the fundamental building block connecting individual functions to complete repositories, remains systematically underexplored. Understanding class-level capabilities is a prerequisite to repository-level success, as current approaches cannot distinguish whether failures stem from basic implementation deficiencies or sophisticated context management issues. We introduce RealClassEval, a benchmark derived from real-world open-source repositories, comprising 400 classes divided into seen and unseen partitions based on model training cutoffs. Our systematic evaluation across seven state-of-the-art LLMs reveals fundamental limitations masked by synthetic benchmarks: while LLMs achieve 84 to 89% correctness on synthetic tasks, they attain only 25 to 34% on real-world classes, with minimal distinction between seen and unseen codebases. Notably, this performance range converges with concurrent work using rich natural language descriptions and repository context, suggesting that specification complexity cannot overcome core implementation bottlenecks. Documentation provides marginal improvements (1 to 3%), while retrieval augmentation yields greater gains (4 to7%), specifically with incomplete specifications. Error analysis reveals AttributeError, TypeError, and AssertionError dominate 84% of failures, with distinct patterns between synthetic (logic-focused) and real-world (object-oriented) scenarios. These findings establish the diagnostic foundation necessary for advancing both fundamental LLM capabilities and sophisticated repository-level approaches, while providing actionable guidance for production deployment strategies.

cs.SE · cs.AI · cs.LG

新论文 · v1arXiv:2611.20132

CapCover Filter: Semantic Approximate Membership Queries over Embeddings

Dvir David Biton, Shay Hahiashvili, Nadav Keren, Gil Einziger, Roy Friedman

原文摘要

Determining whether an embedding is sufficiently similar to any vector in a vector store without performing a nearest-neighbor search can substantially improve query throughput in vector-store systems. We study Semantic Approximate Membership Queries (SAMQs), which allow queries to be rejected when no indexed vector meets the requested similarity threshold, while guaranteeing no false negatives. We introduce CapCover Filter, a geometric data structure that summarizes a set of embeddings using spherical caps. Its representation is independent of the query threshold, allowing a single filter to support arbitrary thresholds without reconstruction. We evaluate alternative construction and quantization methods, compare CapCover Filter with competing filters, and measure its impact on query throughput. Across five real-world datasets, at a cosine similarity threshold of 0.95, CapCover Filter eliminates approximately 20--100\% of index accesses while using only 2\% of the raw (FP32) embedding memory. When placed before a FAISS flat index, it improves query throughput by up to 382\% and by approximately 116\% on average across the five datasets, while returning the same range-search result. We also evaluate lower similarity thresholds. Our results show that CapCover Filter can substantially improve vector-store throughput when similarity thresholds are tight, retrieval is expensive, and indexed embeddings are highly clustered. We release an optimized implementation of CapCover Filter and its evaluation framework.

cs.DB · cs.AI

新论文 · v1arXiv:2611.20129

Can decision models extract clinical data from cancer pathology reports? An evaluation of accuracy, calibration, and failure modes

Indranil Mallick, Paromita Roy

原文摘要

Decision models are a new class of language model that, instead of generating text, choose one option per question from a fixed list and return a probability for every option. They have not yet been evaluated on clinical text. We evaluated two open-weights decision models, Nimble 9B and Tev1 4B, for extracting 25 fields from 1,113 oral cavity cancer pathology reports from a single centre, against a prospectively maintained clinical database. We compared them with two generative models: Qwen3.5 9B, run locally, and GLM-5.3-flash, run through a cloud API. Tev1 reached a mean field accuracy of 0.917 (95% CI 0.913 to 0.920), the same as Qwen3.5 9B (0.918; p = 0.26), and was four times faster (15 s against 61 s per report). Nimble reached 0.862 and GLM-5.3-flash 0.986. Both decision models matched GLM-5.3-flash on simple dichotomous categories. They were weak on measurements binned into categories, on fields whose correct answer was not reported, and on some complex decisions (e.g. involved margins). We found that the main cause of the inferior performance of Nimble's came from asking 25 questions in one request: with two questions per request, its accuracy improved. Pooled over fields, the returned probabilities were well calibrated (expected calibration error 0.022 and 0.026), but they did not flag systematic errors within single fields. A simulated strategy of sending answers with a probability below 0.94 to GLM-5.3-flash raised Tev1's accuracy to 0.982, with 38% of answers escalated. Decision models can extract explicitly stated findings from pathology reports quickly on local hardware and return useful probabilities. They should be validated in the configuration in which they will be used, restricted to the fields in which they are reliable, and combined with other methods for numeric values.

cs.CL · cs.AI · cs.LG

版本更新 · v2arXiv:2611.13015

From Questions to Decisions: Benchmarking How LLM Agents Acquire Decision-Relevant Information from Human Beings

Yuhan Tang, Ke Fan, Xinkai Zou, Baichuan Mo, Kangxin Cui, Ao Qu, Ou Zheng, Jiangbo Yu 等 12 位作者

原文摘要

LLM agents are increasingly asked to make plans for people, such as scheduling drivers or assigning nurses. However, people's requests are often incomplete, and the supporting information for decision-making may be outdated, requiring agents to ask for clarification. Existing benchmarks typically provide fully specified problems or rely on uncalibrated LLM-simulated users, leaving this ability insufficiently tested. We introduce Ask2Decide, a benchmark of 43 planning tasks that evaluates whether agents ask for missing information that matters for their decisions. An operator holds facts absent from the agent's request and records, and optimization solvers determine which facts change the optimal plan. We also include similarly worded facts that leave the optimal plan unchanged to test whether agents distinguish useful information from irrelevant details. We collect 79 interaction sessions from 19 participants and find that, after an incorrect plan, human replies omit the correcting fact far more often than replies from an uncalibrated simulated user (43% versus 7%). Using these interactions, we calibrate a simulated operator that more closely matches human reply patterns. Evaluating four frontier LLMs, we find that their initial plans are infeasible on 58-98% of tasks when information is missing, compared with 5-28% when all facts are provided upfront. Agents uncover facts that change the optimal plan only slightly more often than facts that leave it unchanged (39-52% versus 32-41%). These results show that obtaining decision-relevant information through interaction remains a major bottleneck for LLM planning. This is ongoing work and will be updated. The date of this version is shown on the first page.

cs.AI

版本更新 · v2arXiv:2606.13934

Adversarial Concept Search: Predicting Compositional Errors From Feature Geometry

Jennifer Meng Lu, Ruochen Zhang, Isabelle Lee, David Alvarez-Melis, Ellie Pavlick, Naomi Saphra

原文摘要

Humans cannot always intuit what scenarios are most challenging to LLMs. Hoping to capture challenging edge cases, developers either design problems to be difficult for humans or curate extensive benchmarks. What if we could instead anticipate which scenarios a model will fail on? In this paper, we use an LLM's representational geometry to predict which concept combinations it will fail on. We attribute this compositional failure to interference between salient features. In tasks that require systematic composition - toy programmatic settings, multihop reasoning, multilingual factual recall - we find that when a pair of concepts is encoded near-orthogonally, the model reliably composes them. When their linear encodings are close, producing interference, the model fails to compose them. Our method reliably anticipates failure modes across different compositional tasks, without evaluating specific inputs. These results lay the groundwork to use representational geometry to identify high-risk examples, construct targeted stress tests, and provide a scalable foundation for active learning in real-world deployment.

cs.AI

新论文 · v1arXiv:2611.20119

CECAR: Cache & Expert Co-Aware Routing Accelerates On-Device Inference of MoE LLMs

Byeongju Kim, Hoonki Lee, Byungjun Kim, Inyul Ra, Yongjin Kim, Yongseong Lee, Sangyeob Kim

原文摘要

Despite growing interest in on-device LLM deployment, large-scale Mixture-of-Experts (MoE) models remain impractical on resource-constrained devices, as sparse expert activation still requires all expert weights to be memory-resident. To address this, we identify three key observations for on-device MoE inference: (a) cache hit rates are fundamentally bounded even under Belady's optimal policy; (b) these bounds are insufficient for low-latency on-device inference; and (c) prior cache-aware routing based on binary cache presence can induce routing instability. Based on these insights, we propose CECAR (Cache & Expert Co-Aware Routing), a MoE inference system that jointly considers expert routing and cache management. Unlike prior cache-aware routing methods, CECAR uses an ML-based cache policy to predict the expected reuse distance of each expert and prioritizes routing and execution toward experts with smaller predicted reuse distances. On the Qwen3-30B-A3B model, CECAR achieves 12.5 tokens/s on a consumer-grade GPU with only 16 GB of VRAM, where non-resident experts are fetched from SSD, delivering a 4.5x speedup over conventional MoE inference and exceeding the decode speed of baselines that keep the entire model fully resident in DRAM. Code is available at https://github.com/hoonki2002/CECAR.

cs.LG · cs.AI

新论文 · v1arXiv:2611.20116

Locational Marginal Pricing of Compute: Co-optimizing Compute and Energy for Transparent Data Center Flexibility

Vladimir Pyltsov, Teng Zeng

原文摘要

Demand from data centers and AI compute is growing rapidly and is increasing stress on the power grid. This load could be flexible through user adjustments and provide relief during energy scarcity events, yet today its flexibility is provided through utility call-ups or system-operator load shedding, and is invisible to the users whose workloads are being curtailed. In this work, we propose to quantify and couple the prices of compute and energy in a single optimization problem. The resulting compute price reflects the marginal cost of the binding constraints and decomposes into components indicating which grid, power, hardware, or service requirement causes a price surge. The economic interpretation updates the price with real-time conditions and gives users a signal for potential flexibility provision. The co-optimization ultimately connects compute and energy markets while providing a signal for relief of stress on both the grid and the compute resources.

eess.SY · cs.AI

新论文 · v1arXiv:2611.20111

When Valid Decisions Are Not Enough: Evaluating Structural Constraints in Knowledge-Guided Generative Systems

Josef Pavlíček, Petra Pavlíčková, Martin Molhanec

原文摘要

Knowledge-guided generative systems use explicit domain constraints to improve decision consistency and alignment with human-created solutions. However, locally valid decisions do not necessarily preserve contextual relationships in human decision-making. This paper examines when structural constraints improve or degrade agreement with human reference solutions. We introduce HARMONIX, an experimental framework using automated musical harmonization as a controlled testbed for knowledge-guided decision-making. Four configurations share the same Viterbi-based optimization mechanism: a baseline, repetition constraints, functional constraints, and their combination. Experiments cover 19,745 records and 143,634 motif pairs across training and validation partitions. Combined constraints improve reference root-interval agreement by approximately five percentage points in stable structural contexts but reduce it by nearly nine percentage points in contexts involving transposition tracking. These opposing effects are consistent across both partitions, while overall exact chord agreement changes only marginally. Contexts are identified retrospectively using reference annotations; no adaptive constraint-selection mechanism is evaluated. The findings demonstrate that interpretable constraints can introduce systematic context-specific failures concealed by aggregate accuracy metrics. They highlight the need to evaluate relationships between decisions and the contextual applicability of constraints in knowledge-guided generative systems.

cs.AI · cs.CY

新论文 · v1arXiv:2611.20109

MT-RGCN-DTI: A Multi-Task Relational Graph Convolutional Network for Drug-Target Interaction Prediction

Dheeran Sankaran, R Yadunandan, Daniele Santoni, Nimisha Ghosh

原文摘要

Prediction of drug-target interaction (DTI) has become a very crucial part of drug discovery. With the increase in the data volume, it has become very important to improve prediction accuracy by integrating multiple data sources. In this regard, utilising multiple resources in the form of heterogeneous information network to address DTI prediction can be rewarding. In this work, we propose a heterogeneous information network based multi-task DTI prediction model which leverages the advantages of relation graph convolutional network (RGCN). To achieve this, the model employs drug-target interaction network, drug-drug similarity network and protein-protein similarity network to construct a heterogeneous information network. Moreover, we use a knowledge graph embedding scoring functions to model interactions between drug and targets, specifically DistMult and ComplEx as decoders over the learned node representations. We performed extensive experiments on multiple benchmark datasets and compared with state-of-the-art methods. We observed that the proposed model shows competitive performance, thereby confirming the effectiveness of our approach. The codes and datasets are available at https://github.com/Dheeran-S/MT-RGCN-DTI/.

cs.LG · cs.AI

新论文 · v1arXiv:2611.20107

SAFER: Navigating the Exploration-Exploitation Dilemma in World-Model-Based Planning for Autonomous Driving

Yansong Qu, Zilin Huang, Zihao Sheng, Sikai Chen, Samuel Labi

原文摘要

An autonomous vehicle must foresee the consequences of its actions, especially in the rare, time-critical situations where crashes happen. A world model offers this foresight, but only within the coverage of its training data, which poses an exploration-exploitation dilemma. Within this coverage a planner can hardly improve on the data, and beyond it the model's errors mislead it. Logs of competent driving rarely contain the failures that would widen the coverage, and agents that learn from interaction gain them only by failing themselves, without being shown what to do instead. We propose SAFER, a framework that learns offline from failures that others make, judged by an expert, and keeps learning online from its own driving. Offline, partially trained agents and scripted, recoverable perturbations Stage failures in simulation, and an expert Annotates every visited state with its action and value, so that much of what the planner explores is covered before it drives. In planning, a chunked latent world model Foresees the consequences fast enough for real-time use, and a risk-aware sampling-based planner Evaluates candidate maneuvers by their return minus a bounded collision penalty. Online, a loop continuously Refines the model wherever this driving leaves the coverage of the offline data. In closed-loop evaluation against the compared methods, SAFER reaches the highest route completion and reward per step on the CARLA Leaderboard routes, 0.95 and 0.77 from a privileged bird's-eye view, and the highest episode reward in MetaDrive, 53.4 in regular and 49.0 in safety-critical traffic. Project page: https://ys-qu.github.io/safer-website.

cs.RO · cs.AI

新论文 · v1arXiv:2611.20106

Making Preemption Productive in Long-Output LLM Serving

Ziji Shi, Xinyi Wan, Chaoyi Ruan, Guangxing Huang, Min Lin, Jialin Li

原文摘要

Reasoning LLMs generate long outputs whose key--value (KV) caches occupy GPU memory for minutes. When memory fills, the server preempts requests, and a preempted request can repeatedly lose and recover its state while producing little new output. We call this behavior KV thrashing. In certain cases, existing systems leave more than a fifth of requests stalled for over a minute. Because preemption cannot be avoided without delaying admission, we ask how to make it productive. We present Stint, a serving system that preserves a preempted request's work, bounds how long it waits, and gives it fresh output after it returns. Stint stores fixed-length KV snapshots in a contiguity-first CPU pool with an exact-extent allocator, and moves them with fused K/V gather/scatter kernels and double buffering. Age-based readmission brings stalled requests back, and protection gives them a turn of fresh decoding before another preemption. Remaining-length estimates from decoding hidden states guide victim selection. On olympiad-math replays with Qwen3-32B on H200 GPUs, no request pauses for more than 50 seconds under Stint and p99 time per output token falls by 50.7\% relative to recomputation in the same engine. The replay as a whole finishes 10.7\% sooner (12.0\% higher throughput), because the requests that baselines starve no longer finish last.

cs.DC · cs.AI · cs.OS

新论文 · v1arXiv:2611.20094

FRAAC: A French Multimodal Large Language Model and Dataset forPictogram-Based Augmentative and Alternative Communication

Alexandre Audibert, François Portet, Benoit Favre, Maja Jonck Hjuler

原文摘要

Pictograms are a core modality in augmentative and alternative communication (AAC), yet recent multimodal large language models (MLLMs) cannot reliably interpret or produce pictogram sequences. Two obstacles explain this gap: the absence of a standardized pictogram grammar and the scarcity of parallel text-pictogram data. We introduce Retrieval-Augmented Pictogram Selection (RAPS), an annotation method that refines a rule-based transcription system's output with LLM-based concept extraction and retrieval over a pictogram vocabulary, handling cases where hand-written rules fail (named entities, numbers, polysemous words) without requiring new rules. Applied to French, RAPS yields a synthetic corpus of 1.4 million parallel pairs, of which 300k are publicly released. In a human ranking study where annotators order four pictogram transcriptions of the same sentence, RAPS outperforms the rule-based state of the art. Moreover, we show that standard reference-based metrics separate systems of clearly different quality but fail to rank strong systems against each other. Finally, we build an instruction-tuned model that supports interaction in several AAC modalities: it accepts pictogram or text input and answers in pictograms, speech, or text.

cs.CL · cs.AI

版本更新 · v2arXiv:2611.07940

CardioLM: A multilingual suite of small language models for the cardiology domain

Manuela-Daniela Danu, Miguel Rodríguez-Ortega, Alberto Becerra-Tome, Pablo Arancibia, Fernando Gallego-Donoso, Lorenzo Pratesi, Xi Bai, Andrei-Damian Dinescu 等 40 位作者

原文摘要

Background: Cardiovascular care generates large volumes of clinical free-text documentation in local languages. However, clinical natural language processing (NLP) resources remain concentrated in English, and no existing benchmark is cardiology-specific, multilingual, and broad in entity coverage. Methods: We present two contributions. The first is CardioCCC, a corpus of 508 cardiology case reports originally published in Spanish and extended to Czech, Dutch, English, Italian, Romanian, and Swedish through machine translation, annotation projection, and clinician review, with annotations covering diseases, symptoms, procedures, and medications. The second is CardioLM, comprising CardioBERTa, a set of language-specific encoders obtained by continued pretraining on cardiology and biomedical corpora, and CardioNER, a set of named entity recognition (NER) models fine-tuned on CardioCCC. Results: Cardiology-specific pretraining reduced pseudo-perplexity by up to a factor of eight, yet changed strict span-level F1 by at most 0.04, with benefits concentrated in languages whose backbones lacked biomedical coverage. The jointly trained multilingual XLM-RoBERTa-large models achieved the best strict F1 in nearly every language and entity type, exceeding the strongest monolingual system by up to 0.13 F1. A large share of residual errors concerned boundary detection: relaxed matching recovered 0.10-0.23 F1 for disease, procedure, and symptom, but only 0.01-0.07 for medication. Clinicians rated CardioNER well above a rule- and dictionary-based baseline, with the mean rating across entity types rising from 26% to 73%. Conclusions: Specialty-specific pretraining is not the decisive factor for downstream entity recognition, whereas multilingual joint training yields the strongest results in nearly all settings and a simpler deployment path for multi-country clinical information extraction.

cs.CL · cs.AI

版本更新 · v2arXiv:2603.06950

How Private Are DNA Embeddings? Inverting Foundation Model Representations of Genomic Sequences

Jules Kreuer, Sofiane Ouaari, Nico Pfeifer

原文摘要

Genomic foundation models are increasingly used to share dense representations rather than raw DNA sequences. This is particularly pursued for mean-pooled representations that are later used for downstream prediction tasks. Whether this practice protects the privacy of the underlying sequence is largely unknown. We evaluated model-inversions against zero-shot embeddings from DNABERT-2, Evo 2, and Nucleotide Transformer v2 (NTv2). Substantial reconstruction remains possible: a single multi-length Length-Conditioned Query Decoder reconstructed NTv2 sequences with mean Levenshtein similarity from 0.98 at 10 nucleotides to 0.68 at 100, and Evo 2 sequences from nearly 1.00 at 15 nucleotides to 0.67 at 100. DNABERT-2 showed higher resistance (0.56 to 0.51), although it remained above the random baseline at short lengths. Therefore, we further evaluated, whether a linkage attack similar to Homer's attack is possible based on the mean-pooled representations of human DNA sequences. At SNP-enriched length-25 loci and known coordinates, Evo 2 and NTv2 reached 99\% identification at 32 loci and DNABERT-2 at 256 loci. With alignment-based coordinate recovery, the corresponding 99\% budgets were 64 and 128 loci for Evo 2 and NTv2, while DNABERT-2 did not reach the reported identification thresholds. Random loci required larger budgets. These findings show that mean-pooled embeddings can leak sequence and identity, with risk depending on sequence length, representation, locus informativeness, and coordinate information.

q-bio.GN · cs.AI · cs.LG

版本更新 · v3arXiv:2603.27910

GAAMA: Graph Augmented Associative Memory for Agents

Swarna Kamal Paul, Shubhendu Sharma, Nitin Sareen

原文摘要

AI agents that talk with users across many sessions need long-term memory to stay coherent and personalized. Existing approaches use either flat retrieval-augmented generation (RAG) that ignores how memories connect, or entity-centric knowledge graphs. Entity-centric graphs split each message into separate triples, which loses its context, and in conversations the speakers become hubs linked to almost every message, so propagation through them tells retrieval little. We present GAAMA, a graph-augmented associative memory that organizes conversations around topics. It keeps verbatim episodes, uses an LLM to distill them into atomic facts and topic concepts, and then synthesizes higher-order reflections. These four node types are linked by five kinds of structural edges, and topic concepts supply the cross-cutting paths, so mega-hubs do not form. Retrieval blends nearest-neighbor similarity search with Personalized PageRank whose transitions depend on edge type. We also introduce GRAFT (Graph Repair by Augmenting Facts & Topology), an offline repair step: for questions whose retrieved evidence fell short, a bounded ReAct agent searches memory again and records the grounded answer as a fresh fact, so later questions find it directly without extra query-time cost. On LoCoMo10, GAAMA achieves the highest average score of all compared methods (79.1%) and is competitive in every question category; GRAFT adds a small further gain. In a controlled comparison, replacing GAAMA's concept-linked facts with entityrelation triples lowers reward by 3.0 pp. On MemoryArena, GAAMA also has the best average across the three datasets.

cs.AI · cs.IR · cs.MA