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

+729新增论文
+575其中 AI 相关
+82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 16:02 — 2026-11-21 16:02(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)
版本更新 · v2arXiv:2605.20242

A closed-loop framework for perovskite precursor additive discovery

Xin-De Wang, Zhi-Rui Chen, Ze-Feng Gao, Peng-Jie Guo, Cheng Mu, Zhong-Yi Lu

原文摘要

Efficient discovery of precursor additives is essential for improving the performance of perovskite solar cells, yet the large chemical space makes conventional trial-and-error screening inefficient. Here we show that LEAP (LLM-driven Exploration via Active Learning for Perovskites), a closed-loop framework combining a domain-specialized language model with active learning, prioritizes perovskite precursor additives under low-data conditions. The model extracts mechanism-relevant knowledge from the literature and converts it into interpretable molecular descriptors for uncertainty-aware Bayesian optimization. On an evaluation drawn from unseen literature, the specialized model provides more mechanism-consistent answers than the tested general-purpose models. In an expert-guided proof-of-concept study, two later-round additives increase average device power conversion efficiencies to 20.13% and 20.87%, compared with 19.25% for the control, and the best device reaches 21.32%. Because only a limited number of candidates are tested prospectively, these results provide preliminary evidence that literature-grounded molecular representations, Bayesian optimization, and expert assessment can support additive discovery for perovskite solar cells.

cs.LG · cond-mat.mtrl-sci · cs.AI · physics.chem-ph

新论文 · v1arXiv:2611.22219

Fine-Grained Japanese Clinical Named Entity Recognition: A Controlled Comparison of BERT and Open-Weight Large Language Models

Feng Han, Ayako Yagahara, Kenji Hirata, Atsushi Yoshida, Kohsuke Kudo

原文摘要

Background: Generative large language models (LLMs) offer an alternative to encoder-based clinical named entity recognition (NER), but their performance on fine-grained Japanese extraction tasks and sensitivity to generation settings remain uncertain. Methods: We compared two fine-tuned BERT-CRF baselines with nine open-weight LLMs on iCorpus, which defines 113 entity categories. The evaluation used 433 sentences from 36 test documents; 1,672 sentences were used for training. LLMs were evaluated using zero-shot prompting with and without JSON Schema constraints and three-seed QLoRA fine-tuning. Fine-tuned models were assessed under different output-token limits, including 900 and larger model-specific limits of 2,600 or 4,096 tokens. Performance was measured using sentence-level label-surface matching with relaxed and exact-surface F1. Results: The highest zero-shot relaxed F1 was 0.337, compared with 0.836-0.841 for the BERT baselines. At the larger output limits, all eight fine-tuned LLMs with at least 13B total parameters exceeded both BERT baselines (relaxed F1, 0.882-0.914; exact-surface F1, 0.840-0.880). Across three-seed means, increasing the output limit improved relaxed F1 by 0.052-0.085, alongside higher recall and fewer length-limited outputs. JSON Schema constraints changed relaxed F1 by -0.014 to +0.008 among seven models outside the two reasoning-oriented configurations, while medical-general model pairs showed no consistent advantage for medical specialization. Conclusions: On this historical Japanese case-report benchmark, task-specific fine-tuning combined with adequate generation capacity was associated with stronger LLM extraction performance. Output-token limits materially affected comparisons with BERT, whereas enforcing output structure alone did not overcome low zero-shot accuracy.

cs.AI · cs.CL

新论文 · v1arXiv:2611.22218

Learning Optimal Transport Plans Via Autoregressive Token Regression

Ivan Jaen Marquez, Sotirios Panagiotis Chytas, Vikas Singh

原文摘要

Optimal Transport (OT) provides a principled framework for finding the minimum cost coupling between probability distributions. Classical solvers remain the gold standard for this task. At the same time, OT is an ideal testbed for studying whether learning-based amortized models can internalize the structure of OT type problems. Here, ground-truth solutions are available and marginal constraints are explicitly defined. Also, exact discrete solutions have well-characterized sparsity. We describe ToR-OT, a transformer-based model that learns to predict discrete OT plans through autoregressive token regression. We reframe OT into a structured sequence generation task. ToR-OT decodes sparse transport plans as sequences instead of solving an optimization task at test time. We focus on three properties: (a) Zero-shot generalization: pretraining on a large synthetic corpus of OT instances allows a single model to generalize to unseen distribution pairs and heterogeneous cost metrics without retraining, (b) Sparse primal decoding: by exploiting the inherent sparsity in exact discrete OT solutions, the decoder predicts only nonzero entries (avoids dense coupling outputs), (c) Amortized inference: on low-dimensional discrete problems, ToR-OT gives accurate approximate plans and compares favorably with per-instance neural OT baselines. Overall, ToR-OT is not yet a replacement for classical solvers, but a step toward understanding how autoregressive transformers can learn structure of constrained optimization problems like OT.

cs.LG · cs.AI

新论文 · v1arXiv:2611.22215

Communicate the Change: Efficient Expert Parallelism via Activation Compression

Fengyuan Liu, Jay Gala, Siva Reddy, Dzmitry Bahdanau

原文摘要

Scaling has driven the success of large language models (LLM), with sparse mixture-of-experts (MoE) architectures enabling greater parameter capacity without a proportional increase in computation. At large scales, MoE models rely on expert parallelism to distribute experts across devices. Each MoE layer requires two synchronous collective communication steps during inference: dispatching hidden states to the activated experts and combining the resulting expert outputs. As compute throughput has outpaced interconnect bandwidth, these synchronous collectives have become a major bottleneck to MoE scaling and efficiency. We address this challenge with EchoMoE, a lightweight architecture modification that retrofits existing MoE models, requiring approximately 8,500x fewer training tokens than full pretraining. Motivated by the observation that transformer blocks update the residual stream incrementally, we use a hidden state from an earlier layer as a reference and communicate only the compressed difference between it and the current state, rather than communicating each hidden state in full. EchoMoE periodically refreshes these reference states and transmits only the compressed differences synchronously, reducing the traffic on latency-critical paths without significant computational overhead. EchoMoE reduces inter-node communication volume by 58% and improves throughput by 22.6%, while preserving average performance within 3.8 percentage points across 18 benchmarks. EchoMoE further improves inference throughput by up to 2.38x over bandwidth-constrained Ethernet, paving the way for wide-area distributed inference. EchoMoE also tolerates quantization of the communicated representations without measurable quality loss. We release an efficient implementation for applying EchoMoE to pretrained MoE models, together with an inference integration for vLLM.

cs.LG · cs.AI · cs.DC

版本更新 · v2arXiv:2611.20140

Can AI Judge You? An Empirical Audit of Deployment Readiness in LLM-assisted Hiring Systems

Spandan Roy, Subhabrata Majumdar, Soudeep Deb

原文摘要

Delegating resume screening to large language models (LLMs) requires organizations to establish whether they can distinguish candidates by relevant qualifications while allowing only authorized users to change screening rules or approve exceptions. To this end, we develop a system architecture that separates candidate evidence, evaluation, permission controls, and assessment record, and empirically examine two complementary dimensions of its deployment readiness. First, we assess its operational readiness on four LLMs using a controlled experiment on 210 resumes across 14 job roles and six decision interfaces. Second, we assess its compliance readiness by testing how it handles unauthorized user requests and instructions embedded in resumes. We find that operational performance depends on the choice of evaluation interface: scalar scores generally distinguish qualification changes better than simple pass/fail judgments, while comparative judgments are sensitive to presentation order. Explicit permission rules and separation of resume content from instructions reduce unauthorized compliance, to a varying degree of effectiveness across LLMs and usage contexts. Finally, we find that LLM evaluators often diverge between their decisions and offered justifications. Going beyond simply evaluating LLMs for hiring tasks, our framework supports an audit of a fully configured resume evaluation system to enable its deployment decision.

cs.CY · cs.AI · cs.CL

版本更新 · v3arXiv:2604.00024

WHBench: Evaluating Frontier LLMs with Expert-in-the-Loop Validation on Women's Health Topics

Sneha Maurya, Spandana Govindgari, Girish Kumar

原文摘要

Large language models are increasingly used for medical guidance, but women's health remains under-evaluated in benchmark design. We present the Women's Health Benchmark (WHBench), a targeted evaluation suite of 47 expert-crafted scenarios across 10 women's health topics, designed to expose clinically meaningful failure modes including outdated guidelines, unsafe omissions, dosing errors, and equity-related blind spots. We evaluate 22 models using a 23-criterion rubric spanning clinical accuracy, completeness, safety, communication quality, instruction following, equity, uncertainty handling, and guideline adherence, with safety-weighted penalties and server-side score recalculation. Across 3,102 attempted responses (3,100 scored), no model mean performance exceeds 75 percent; the best model reaches 72.1 percent. Even top models show low fully correct rates and substantial variation in harm rates. Inter-rater reliability is moderate at the response label level but high for model ranking, supporting WHBench utility for comparative system evaluation while highlighting the need for expert oversight in clinical deployment. WHBench provides a public, failure-mode-aware benchmark to track safer and more equitable progress in womens health AI.

cs.CL · cs.AI · cs.CY

版本更新 · v2arXiv:2608.12387

Query Timing Produces Opposite Positional Biases Between LLMs and Humans

Jasin Cekinmez, Addison J. Wu, Thomas L. Griffiths

原文摘要

Positional biases such as recency and primacy effects have been documented in large language models (LLMs), yet the underlying mechanism by which these models make their evaluations remains poorly understood. Both primacy and recency biases have been observed in human judgments in response to evidence, but recent work suggest that when the listener updates their beliefs - during the presentation of evidence or only at the end - influences the presence of such effects. We investigate whether a similar phenomenon holds for LLMs, finding divergence from human behavior. These biases are more pronounced in newer models compared to their predecessors.

cs.CL · cs.AI

新论文 · v1arXiv:2611.22203

EGS: A Complementary Safety Framework for Educational LLMs

Zexuan Fei, Zhengnan Guo, Yiming Luo, Kaiyi Huang, Aimin Zhou, Bo Jiang, Fei Tan

原文摘要

Large Language Models (LLMs) offer strong potential for personalized education, but adolescent-facing deployment raises risks of toxic, biased, misleading, or developmentally inappropriate content. We argue that educational safety should move beyond binary refusal: student queries are often ambiguous or curiosity-driven, requiring contextual guidance rather than simple blocking. We introduce the Educational Guardrail System (EGS), a lightweight, model-agnostic runtime framework that complements intrinsic alignment through query moderation, scenario-aware pedagogical guidance, and response moderation. EGS triages inputs by risk level, redirects gray-zone queries with category-specific educational interventions, and audits outputs before delivery. Experiments across general and education-specific benchmarks show that EGS improves moderation performance, refusal quality, and jailbreak robustness across diverse LLMs. These findings suggest that defense-in-depth guardrails can strengthen safety while preserving the pedagogical utility required in educational LLMs. Our code and model weights will be publicly available to support community adoption.

cs.CL · cs.AI

版本更新 · v2arXiv:2608.13087

Sampling Luck Masquerades as Allocation Gain: Auditing Test-Time Budget Allocation for Neural Combinatorial Optimization

Jinhyung Bae

原文摘要

Neural combinatorial optimization (NCO) solvers report the best of many sampled solutions per instance, and the sample count is, by convention, identical for every instance. Whether a non-uniform allocation of a fixed total budget would buy anything has not been measured. We measure it, and we audit the measurement itself. First, on in-distribution workloads the allocation headroom is not detectable. Across three pretrained solvers (POMO, AM, SymNCO) on uniform TSP-100, an oracle allocation computed and evaluated on the same stored samples reports a 2.2-2.6% gain with intervals excluding zero; measured out of sample the same gain is indistinguishable from zero (0.457, 0.015, -0.512 percent). Following the customary in-sample procedure, all three solvers would have supported a published 2%-level gain that does not exist. We calibrate this bias against an instance-wise null in which the true gain is zero by construction; over the ranges we test it does not shrink with more samples or more instances. Second, the same correction that removes the phantom gains preserves a real one. Under distribution shift (a workload mixing uniform and clustered instances), a pre-registered confirmatory experiment finds that allocation guided by held-out sample statistics improves best-of-k by 11.5% (AM, primary endpoint; 95% CI [7.4, 19.7]) and 12.0% (SymNCO, replication) at equal evaluation budget, with the signal-acquisition cost not charged; a pre-registered negative control (POMO, an order of magnitude more robust to shift) shows -0.3% [-0.7, 0.24]. The gain exceeds a frozen distribution-label baseline by 4.2 points [1.9, 7.7]. An exploratory policy charging a 20-sample probe against the same budget retains 3.4% (AM) and 4.6% (SymNCO). We give a correction procedure and a reporting checklist, and release all data, code, and the pre-registration record.

cs.LG · cs.AI · math.OC

版本更新 · v3arXiv:2603.01608

Evaluating and Understanding Scheming Propensity in LLM Agents

Mia Hopman, Jannes Elstner, Maria Avramidou, Amritanshu Prasad, David Lindner

原文摘要

As frontier language models are increasingly deployed as autonomous agents pursuing complex, long-term objectives, there is increased risk of scheming: agents covertly pursuing misaligned goals. Prior work has focused on showing agents are capable of scheming, but their propensity to scheme in realistic scenarios remains underexplored. To understand when agents scheme, we decompose scheming incentives into agent factors and environmental factors. We develop realistic settings allowing us to systematically vary these factors, each with scheming opportunities for agents that pursue instrumentally convergent goals such as self-preservation, resource acquisition, and goal-guarding. Across 11 frontier models and four scenarios, we find almost no scheming despite high environmental incentives, and show this is unlikely to be explained by verbalized evaluation awareness. While inserting adversarially-designed prompt snippets that encourage agency and goal-directedness into an agent's system prompt can induce high scheming rates, snippets used in real agent scaffolds rarely do. In model organisms built with these snippets, scheming behavior is brittle: removing a single tool can drop the scheming rate from 59% to 3%. Increasing oversight can raise rather than deter scheming by up to 25%, an effect that depends on the wording of the oversight sentence and does not appear in a second scenario, where the sabotage itself is reviewed. These results suggest that scheming propensity should be reported per agent configuration rather than as a single rate. Our incentive decomposition enables such systematic measurement in settings relevant for deployment, which is necessary as agents are entrusted with increasingly consequential tasks.

cs.AI

版本更新 · v3arXiv:2606.11722

ICA Lens: Non-Gaussianity Reveals Interpretable Directions in Language Models

Sida Liu, Feijiang Han, Sarah Wiegreffe, Furong Huang

原文摘要

Finding human-interpretable directions in language-model activations is important for understanding and controlling model behavior. These directions are usually discovered through indirect learning objectives such as sparse reconstruction, but it remains unclear what existing statistical structure distinguishes interpretable directions and whether that structure can be used directly for feature discovery. We find that non-Gaussianity provides such a signature: projection distributions of natural-text residual-stream activations along sparse autoencoder (SAE) decoder directions are substantially more non-Gaussian than those along random directions, with most contexts forming a concentrated background and recurring patterns often occupying one tail. This raises a natural question: if interpretable directions tend to have non-Gaussian projection distributions, can we discover them by searching for non-Gaussian directions directly in activation space, without training another sparse dictionary? To test this hypothesis, we introduce ICA Lens, a practical workflow for efficient discovery, reliable annotation, and faithful evaluation of non-Gaussian directions in LLM residual-stream activations. Across GPT-2 Small, Gemma 2 2B, and Qwen 3.5 9B Base, we find that (1) the recovered non-Gaussian directions capture coherent lexical, syntactic, and semantic patterns under both human and LLM-based evaluation; (2) non-Gaussianity is positively associated with both human-label confidence and automated interpretability; and (3) the resulting coordinates are competitive with SAEs across interpretability, sparse reconstruction, concept prediction, and causal interventions, while fitting in minutes per layer and using 7-43 times less memory. These results show that non-Gaussianity is a direct unsupervised signal for discovering interpretable and useful structure in activations.

cs.LG · cs.AI · cs.CL

新论文 · v1arXiv:2611.22188

From Imperfective Paradox to Moorean Absurdity: Single-Chain Reasoning Lock-In and Multi-Candidate Verbalization Across Aspectual and Pragmatic Ambiguity

An-Sheng Jhang

原文摘要

Natural language entailment and coherence judgments often depend on whether an event merely occurred, reached a telic culmination, satisfied a plan, or was evaluated under multiple pragmatic standards. We test observable consequences of the single-chain premature lock-in hypothesis across four preregistered, sealed formal experiments on Qwen3-8B-Q8_0 (16,640 formal units, temperature=0.0): 15 of 16 hypothesis slots retain confirmatory status (12 rejecting H0 under within-experiment Holm control at alpha_fam=0.05; E3_H3 demoted to descriptive status). In Experiment 1, Direct-CoT vs. MCU differs by 25.00 pp on telic high-confidence completion (Holm p=1.31239e-05), while NLI-CoT collapses by -91.25 pp on uninterrupted atelic items (Holm p=1.81805e-21). Multi-Candidate Verbalization (MCV) elicits K=5 weighted estimates in one call, separating the prompt from unweighted (MCU) and self-weighted (MCW) readouts. Mandarin (E2) and English (E3) expose an Interruption-Anchored Realis Paradox under NLI-CoT. In Experiment 4, MCU-CF raises perceived coherence on implicit Moorean sentences by 26.64 pp over Direct-CoT (Holm p=0.0006), alongside Direct-CF - Direct-CoT = 29.15 pp and MCU-CoT - Direct-CoT = 17.31 pp. A 1,800-call OpenAI ablation and a stage="dev" P16.7 cross-model extension (4,240 planned/collected units across five external models: 4,222 valid, 18 format-missing, 0 transport-missing; <2% missing per slice; OpenAI gpt-4o-mini at top_logprobs=20 achieves 1,000/1,000 valid units) show that discrete NLI-CoT and scalar Direct-CoT judgments dissociate across models. Confirmatory effects concern observable prompt-and-readout protocols; internal premature lock-in remains a hypothesis. All code, stimuli, preregistered analysis plans, and verification tools are available at https://github.com/asj252/llm-judgment-studies.

cs.CL · cs.AI

新论文 · v1arXiv:2611.22180

CueClash: Frontier Vision-Language Models Judge 3-D Relations Close to the Picture Plane

Soham Dan

原文摘要

A vision-language model (VLM) can judge 3-D relations from the picture or from the ground layout. In photographs the two usually agree. We separate them with CueClash, rendered scenes in which the 2-D cue for each of five relations is right on exactly half the statements. The shortcut gap is accuracy when the cue is right minus accuracy when it is wrong. For Claude Opus 5.5, GPT-5.5 and Gemini 3.6 Flash at low reasoning effort it is +18 to +37 percentage points. It is +7 to +29 without two relations whose cue-wrong items are confounded with borderline cases or occlusion. Given ground coordinates and each relation's definition, they answer 99 to 100% correctly. Pixel measurements with camera depths remove most of the gap for objects fitting through openings. The failure lies in the step from the picture to the layout. The maps that the models draw of the scene from above mostly keep the depth order but, like the picture, shorten depth. GPT-5.5's maps stay close to a copy of the picture. Opus 5.5's and Gemini's maps undo about 0.6 of this shortening. Gemini's then shrink all depths by a further uniform factor. Opus 5.5 follows the picture mainly on fits, and its maps correct depth about as much as human observers of real scenes do. The three models also lose accuracy on CV-Bench photographs where a reliable 2-D cue misleads. In two open models the layout information reaches the language model, but their maps lose much of it.

cs.CV · cs.AI · cs.RO

新论文 · v1arXiv:2611.22179

RAG in a Trenchcoat: When Minimal Memory Is Enough for Agentic Systems, and When It Isn't

Jingyu Liu, Zongze Li, Zach Xu, Zhanhui Zhou, Tahseen Rabbani, Dawn Song, Ce Zhang

原文摘要

Agentic memory research has accumulated extraction, supersession, reflection, and graph synthesis on top of retrieval-augmented generation. However, we found that almost none of those primitives earn their cost in our tested memory benchmarks: we show this with BRIDGE, a deliberately minimal RAG memory with chained observation notes, hybrid BM25 with cosine retrieval, and chronological rendering, which reaches parity within bootstrap CI with the prior SOTA on LongMemEval-S and BEAM, and beats the best memory baseline on EverMemBench by 5.7-6.3 points at matched reader. It runs at ~5-7K reader-prompt tokens per query, using a single LLM call for each ingestion or query. We argue that this parity is itself the result: standard benchmarks score what memory retrieves, not what memory enables when the agent acts. To test the latter, we introduce five adversarial probes (Rule Application +-, Pattern Induction Style / Procedural, Dynamic Multiturn) judged at generation time by a cross vendor verification chain, where each query can only be solved by applying a remembered rule or pattern it never mentions. Across 5 probes x 20 scenarios, no architecture wins universally: atomic fact extractors (LangMem) win rule application, while BRIDGE's block-summary memory wins pattern induction. This result holds across four domains, three graders, and three reader families. Each design fails where the other wins, for reasons a recall benchmark cannot see: atomic-fact extractors discard the form a demonstrated pattern takes, while block-summary memory retrieves the rule but the reader does not act on it, a gap that presenting the rule more prominently largely closes. What separates memory designs is not what they retrieve but what they enable an agent to do, and that is what agentic memory should be evaluated on.

cs.AI

版本更新 · v3arXiv:2608.00023

Role Steering of Language Models for Social Simulations

Isaac Song, Mohammed Rehan Parwani, Glenn Matlin, Emile Anand, Akhil Theerthala, Arjun Chatterjee, Anthony Wen-Ming Zang, Maria Kostylew 等 11 位作者

原文摘要

Social simulations built from language-model agents need role-conditioned behavior that can be checked before agents are placed into a simulated population. We introduce an activation-steering screening workflow for role-conditioned agents: define a role profile, extract a role-specific direction, sweep four steering coefficients, evaluate role-profile alignment, and pass or flag each candidate configuration. On OLMo-3-7B-Instruct, we apply the workflow to a mixed 275-role inventory with 228 role-agnostic questions, GPT-4.1-mini prompted role references, and GPT-4.1-mini judges. We also introduce CastVectors, which receive higher judged role-profile alignment than an assistant-axis directional control from prior persona-vector work, with mean overall scores of 63.2 versus 41.1 across the tested grid. They also preserve high lexical diversity, while the control drops sharply at larger coefficients. The role-level screen is the main contribution: most roles improve as steering increases, but 38 roles decline across all six measured dimensions, showing why simulation builders should choose coefficients per role rather than deploy a uniform high-strength setting. Our code is available at https://github.com/HCAI-Lab-GT/casting-call-vectors, and the role vectors and evaluation data at https://huggingface.co/datasets/HCAI-Lab-GT/casting-call-vectors.

cs.CL · cs.AI

新论文 · v1arXiv:2611.22178

CorrectCast: Selective Multimodal Context for Scale-Aware Residual Correction in Precipitation Nowcasting

Henan Wang, Shengwu Xiong, Yifang Zhang, Pengfei Duan

原文摘要

Precipitation nowcasting models suffer from errors in intensity, location, and spatial structure. Addressing these errors by retraining nowcaster backbones can be computationally expensive and requires architecture-specific modifications. To overcome these limitations, we propose CorrectCast, a decoupled residual correction framework that refines completed nowcasts while keeping the base nowcaster frozen. By operating on predicted precipitation fields, CorrectCast requires training only a lightweight corrector, reducing computational overhead while providing a unified correction interface across different nowcasting models. Conditioned on the base nowcast and historical quantitative precipitation estimation (QPE) and precipitable water vapor (PWV) observations, the corrector selects multimodal context and employs scale-aware modulation to estimate residual corrections. On the Mississippi River Basin (MRB) and contiguous United States (CONUS) datasets, CorrectCast improves six base nowcasters spanning extrapolation-based, deterministic, and generative approaches, with absolute gains in mean CSI averaging 0.0233 and 0.0290 across the six models, respectively.

cs.AI

新论文 · v1arXiv:2611.22177

The Only Edge Was Memorized in Training: LLM Trading Agents Around Earnings and Bad News

Jackson Wang

原文摘要

Large language model (LLM) trading agents are tested on news, but news reaches stock prices within minutes, and a benchmark that replays headlines at their publication time finds no alpha in trading on them. Money can then remain in two places, before news whose date is known, such as earnings announcements, and after bad news, if the market's fall went too far. However, studies of LLMs and earnings read documents published with or after the announcement, and studies of reversal sort monthly portfolios, so neither tells whether an agent can trade these moments after costs. In this paper, we test LLM agents at both moments on the same look-ahead-free benchmark, before 598 earnings announcements and after 409 bad-news drops of large US stocks. Every input is cut at the decision close, staged agents are compared with single calls, simple rules and a logistic regression, and every decision is made with and without the stock's name. Across five models from four vendors, no agent beat the simple bets. Before earnings, the agents mostly bought and earned what buying before every announcement earned, and after bad news, a Gemini-based agent called almost every fall an overreaction while no agent picked better dips than buying them all. The only gains came from what the models had memorized in training. Given only the ticker and the dates, a GPT-based model called the direction of 59.2% of the announcements before its training cutoff and 53.3% after it. Agent builders should therefore compare an event agent with the simple bet and test it blind after the training cutoff.

cs.AI · cs.MA

版本更新 · v3arXiv:2511.19218

From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment

Xurui Li, Kaisong Song

原文摘要

Large Language Models (LLMs) increasingly power web-facing applications, including conversational assistants and content services, where open user interaction exposes them to diverse safety threats and adaptive attacks. In particular, jailbreak attackers repeatedly reformulate harmful requests to bypass safety alignment. Although existing studies have concentrated on defending dynamic jailbreaks, they often retain discovered attacks as flat samples and discard information about how those attacks were found. In this paper, we introduce MACE, an MCTS-guided adversarial co-evolution framework that turns search into policy-learning signals for both an attacker and a defender. Its search component, GS-MCTS, explores multi-step compositions of rewriting operators and associates discovered attacks with empirical success frequency and path depth. Its optimization component, AC-TGPO, uses these statistics for difficulty-aware group-relative credit assignment and constructs a temporally asymmetric curriculum from current interactions, pairings between the latest attacker and preceding defender, and attacks that survive a defense update. Thus, MACE connects search structure to both within-round policy optimization and cross-round memory, rather than using MCTS only to generate attacks or treating each round as a flat adversarial dataset. Removing difficulty-aware weighting, asymmetric samples, or persistent hard samples increases defense ASR-LR by 3.7 %, 6.0 %, and 5.0 %, respectively. Across evaluated models, the resulting attacker achieves the highest ASR-LR in all six model--dataset settings, while the defender reduces average ASR-LR from 11.9 % for the best baseline to 8.8 %.

cs.CR · cs.AI

新论文 · v1arXiv:2611.22174

Reconfiguring Procedural Knowledge for Compositional Robot Skill Adaptation

Daehee Lee, Dongsu Lee, Sanghyun Ahn, Minjong Yoo, Honguk Woo

原文摘要

General-purpose behavior models increasingly internalize broad procedural knowledge during training, making their adaptation to new goals and skill interfaces less a problem of learning entirely new behaviors than of reconfiguring existing competence. We study this capability as reconfigurability, the ability to realize new behaviors by cost-efficiently restructuring existing components and relations under sparse governing information, where task specifications provide high-level goals and procedure specifications expose a limited set of reusable factors, leaving the concrete bindings and compositions required for execution underdetermined. We propose RecPo, a framework that represents skills task-agnostically through multiple procedure views and materializes executable skill sequences by composing stored transitions to realize task-specific changes. RecPo maintains a procedure knowledge base of skill-induced transitions, with each view exposing a reusable factor and preserving local skill connectivity. At deployment, it interprets task specifications through the same views and resolves the resulting compositional queries over the stored relations to bind skills for execution. On long-horizon tasks in an extended Franka Kitchen benchmark, RecPo achieves 91.9% ordered completion with 1.0 ms of task inference per episode across varying procedure support and task-record sparsity. Real-robot studies further demonstrate adaptation through local memory edits without task-specific retraining.

cs.AI

新论文 · v1arXiv:2611.22171

Readie: Low-Latency Code Sandboxes for AI Agents

Arjun Sivaraman, Zhaoheng Li, Alan Shi, Sanjay Malisetty, Eugene Wu, Yongjoo Park

原文摘要

From quantitative research to personal assistance, AI agents act by iteratively generating and executing code. To decouple the agent harness from code execution, many agents offload execution to dedicated systems called code sandboxes. However, existing code sandboxes suffer from slow startup. Whether they rely on custom snapshots or a single fat image, the root cause is the same: it is hard to build a compact, pre-warmed snapshot that can be shared across diverse agent-generated code while covering nearly all the artifacts (e.g., packages, datasets) it requires. In this work, we propose learned startup snapshots for low-latency code sandboxes. Our insight is that while it is hard for a single snapshot to be both compact and cover most workloads, a set of snapshots can achieve both. Each snapshot is small, initializes quickly, and covers a fraction of agent-generated code; together, the set covers most of it. We learn this set from a large training workload by formulating snapshot construction as an optimization problem and solving it with an approximation algorithm. Planned on 15K public and LLM-synthesized scripts, our system, Readie, reduces startup latency by 3-22$\times$ and end-to-end runtime by 1.3-3.3$\times$ compared to existing approaches at the same storage budget.

cs.AI · cs.DB · cs.SE

新论文 · v1arXiv:2611.22169

Not All Evidence Is Admissible: Evidence Governance for Tool-Augmented Medical Agents

Runxun Zhang, Onofrio Antonio Catalano, Jinqiao Wang, Jingwei Wei

原文摘要

Tool-augmented medical agents treat whatever their tools return as evidence. We show that this default is costly: medical tools fail silently, and a frozen reasoner cannot ignore what enters its context, so many cases it would have answered correctly are flipped to wrong once tool outputs are added. We call this evidence hijacking. $\mathcal{V}$-information explains the phenomenon: the reasoner has lost the option to ignore, so a tool output can carry negative usable information, and moving the decision of what to keep outside the reasoner restores that option. The theory further shows that part of any filtering gain comes merely from letting the reasoner see less, and only selection by content lets it see better. Building on this, we recast tool-augmented reasoning as evidence governance, which separates deciding what counts as evidence from the reasoner that answers, and propose context-independent evidence governance (CIEG), the first medical agent framework that rules, after acquisition, on whether each individual tool output may enter as evidence. Every tool output is written as a card that can be checked against the raw input; an independent checker that sees only the input and the cards decides which cards become evidence; and the reasoner answers from the admitted cards alone, with no training anywhere. On medical question answering, fundus glaucoma screening, and chest X-ray pneumonia classification, it achieves the highest accuracy among models of comparable scale, and significantly outperforms state-of-the-art tool-using agents and multi-agent systems built on the same backbone at lower cost.

cs.AI · cs.CV

版本更新 · v3arXiv:2611.09670

Goodhart in the Planner: Sampling-Based Control Exploits Rewards Learned from Human Preferences

Soovadeep Bakshi, Apurba Bose

原文摘要

Robots increasingly plan by searching against a reward model learned from human preferences. We measure what this search does to the true objective. On HalfCheetah and Hopper with crowd-sourced labels, and on Meta-World drawer-open and plate-slide with synthetic and blind human labels, harder MPPI search cuts true return by up to 76% while held-out label accuracy stays at 0.75-0.97; with human labels the drawer opens in 0 of 20 runs, against 10 of 10 for the cloned policy the planner starts from. The effect persists with a learned dynamics model and an image scorer. We prove that $χ^2$-regularized MPPI at a temperature set by the scorer's error spread on the proposals is worst-case optimal, and that no gain can be certified once that spread exceeds the scorer's own. Calibrated from 100 ranked proposal pairs, this rule is the safest fix we test, keeping 86-94% of the cloned policy's return on drawer-open and plate-slide.

cs.RO · cs.AI · cs.LG

版本更新 · v4arXiv:2608.10450

Persistent Recursive Worlds Enable Autonomous Software Evolution

Beichen Huang, Zhenyu Liang, Bowen Zheng, Ran Cheng

原文摘要

Software systems accumulate dependencies and design commitments over time, while the agents that modify them operate within bounded sessions. How, then, can development remain cumulative across generations of transient agents? We introduce EvoX Genesis, which organizes development as a persistent recursive software world. Agents enter an accepted repository version, delegate work within local scopes, and submit changes for verification and parent review. Once accepted, those changes pass to later episodes together with the contracts and tests used to assess them. The project thus preserves both earlier work and a basis for evaluating what comes next. We examine this continuity through four studies of formation, continuation and redevelopment. Starting without compiler code, Genesis built a C compiler with 219,676 lines of Rust source through 1,019 agent episodes over 123.4 hours. Within this growing implementation, repairs increased passes on a fixed set of 100 Csmith programs from 61 to 93; a subsequent revision retained all 93 passes. Development also continued in a separate compiler lineage when DeepSeek V4 Flash inherited a GLM 5.2 snapshot. To examine continuity when the implementation itself changes, Genesis redeveloped selected MESA functionality in Rust, with numerical agreement assessed against the Fortran reference across six workloads. The WebGL trajectory, meanwhile, exposed both repairs and regressions as repeated checks recorded progress towards 2,017/2,019 conformance pages passed. A complementary EvoX Compiler case connects program-merging performance to executable graph and backend contracts. Together, these records trace how accepted work becomes a foundation for further development: later agents inherit both usable implementations and concrete ways to assess further revisions.

cs.SE · cs.AI · cs.MA · cs.NE

新论文 · v1arXiv:2611.22161

What RL on Expert-Built Tasks Teaches Nemotron 3.5 Super

Sushant Mehta, Edwin Chen

原文摘要

We post-train NVIDIA's Nemotron 3.5 Super (120B total, 12B active parameters) with reinforcement learning (RL) on five expert-built task datasets, one at a time: real-world software engineering, complex constraints, long-horizon tool use, professional chart understanding and enterprise document reasoning. Each run trains a fresh rank-32 LoRA adapter with CISPO on Tinker, with no supervised stage, and each adapter is compared with the base model on the same tasks, harness and settings. On twelve headline benchmarks, eight built by other groups and four by Surge AI, the gains range from +4.5 points on IFBench to +13.8 on GDP_pdf, which rises from 10.4% to 24.2%; SWE-bench Pro V2 rises from 86.9% to 92.1%, Terminal-Bench 2.1 from 63.9% to 70.3%, MultiChallenge from 58.3% to 66.6%, AutomationBench from 14.8% to 27.5% and Chartography from 39.4% to 50.6%. The document adapter, trained without API tools, also gains +8.0 points on AutomationBench. In attempts read side by side with the base model's on tasks each adapter newly solves, the base model's failures recur in a few modes, and each adapter does more often what the base model missed: the tool-use adapter finds the data and the rules before acting, the coding adapter writes its own test or scratch files on SWE-bench Pro V2, the constraints adapter carries earlier turns into its answers, the chart adapter measures the marks in code, and the document adapter reads AutomationBench's spreadsheets.

cs.LG · cs.AI · cs.CL · cs.SE

版本更新 · v3arXiv:2610.14350

Natural Code Transformation Confuses Agents in Vulnerability Analysis

Jie Zhu, Ronik Bhaskar, Heather Zheng, Ben Y. Zhao, Kexin Pei

原文摘要

Having demonstrated a strong potential for automating vulnerability discovery, AI agents can also be used to attack software systems at scale. Today's software systems need mechanisms to level the playing field by slowing down potential attacks, giving them time to patch vulnerabilities and strengthen software. Along this line, our study of agent trajectories in vulnerability analysis shows a surprising finding: agents produce dramatically different results across real-world code commits that carry the same functionality and the same vulnerability. This phenomenon appears consistently in multiple code domains and vulnerability categories. It motivates us to study natural and functionality-preserving code transformations derived from real-world software development, and their potential as a defense against attacker agents. We build a corpus of natural transformations by mining recurring development patterns from real-world code commits, and apply them to evaluate impact on agentic vulnerability analysis. Our results show that even a single form of transformation can effectively "confuse" frontier agents, e.g., Claude Mythos 5.2 with Claude Code, reducing successes in vulnerability discovery. We also study different agent failures, whether transformations can be composed to provide stronger protection, and adaptive attackers who seek to reverse them. Overall, our study of natural code transformation significantly expands the research space of using code transformations to defend against malicious agents.

cs.CR · cs.AI · cs.LG

新论文 · v1arXiv:2611.22151

EgoHandWAM: Probing the Role of Hand Motion in Egocentric World Action Modeling

Kailin Wang, Haoxiang Jie, Yaoyuan Yan, Zhiyou Heng, Zhaosong Li

原文摘要

How much does hand motion tell us about what a first-person camera will see next? EgoHandWAM forecasts three-dimensional wrist and fingertip trajectories and a future visual feature from the current image and recent hand states. We train on 751,065 EgoDex windows using a frozen DINOv2 encoder. On held-out development sessions, the action-only model reduces two-second wrist endpoint error from 60.26 to 56.48~mm relative to a stationary forecast. It still underestimates long reaches. Adding future-feature supervision gives nearly identical trajectory accuracy. With the recorded hand history, the visual head reduces smooth L1 error by 1.32\% relative to copying the current feature. Shuffling histories across sessions erases this gain. In a separate four-task probe, adding recorded future displacements to observed context and future tracking-validity masks reduces visual-feature error by 3.93\%. Masks alone give little improvement. A trajectory predicted from the same observations adds little. On two simulated RoboTwin tasks, substituting aligned human examples for an auxiliary batch of robot examples gives no consistent increase in task success across three training seeds.

cs.CV · cs.AI

新论文 · v1arXiv:2611.22148

Long-Context Generation Is a Sampling Problem (When You Turn Up the Heat)

Allen G. Roush, Minh Nhat Nguyen, Ravid Shwartz-Ziv, Judah Goldfeder, Sanjay Basu

原文摘要

Open-weight language models are widely observed to get worse the longer they generate. Often, attempts to fix this involve more training. At high temperature, The sampler, i.e. the rule that picks the next token out of the model's probabilities, decides whether long and agentic generation survives and/or thrives. We test 29 open-weight models from 1B to 122B parameters on ten established benchmarks plus needle retrieval out to one million tokens, covering SWE-bench Verified coding agents, long chain-of-thought reasoning, long-form writing, long structured output and long-context reading. Samplers that keep their pool of candidate tokens small and steady as temperature rises don't lose accuracy. Samplers that let in low-probability tokens fall apart. At temperature 2, coding agents running top-p, top-k or Top-H resolve 0 of 100 GitHub issues, while the same agents running top-nsigma resolve 46 to 62 and win every issue the two samplers disagree on, on three Qwen3.5 models. Over all benchmarks top-nsigma keeps a median 100% of its temperature-1 score where top-p keeps 26%. Counting how many tokens each sampler lets through explains why. In one Llama-3.1-8B story, going from temperature 1 to 2 grows top-p's pool from 4 tokens to 51,845 while top-nsigma stays at 3. Across 133 combinations of sampler setting and temperature, the typical pool size predicts judged quality (Spearman correlation of -0.77), and when we pick the pool at temperature 1 and only apply temperature afterwards, Top-H, min-p and p-less stop collapsing. At the default temperature of 1 most samplers perform about the same, We release all code and data to the public.

cs.AI · cs.CL

新论文 · v1arXiv:2611.22144

When does deep reinforcement learning pay off in HVAC supervisory control? Four small Modelica case studies trained directly on FMUs

Bingnan Zhao, Harshala Gammulle, Tharindu Fernando

原文摘要

Deep reinforcement learning (DRL) is often proposed for the supervisory control of heating, ventilation and air-conditioning (HVAC) systems. The reported gains are, however, difficult to interpret, because the baselines are often weak and the attainable optimum is seldom known. This paper examines four small, reproducible case studies based on an office building in Brisbane, Australia: pre-cooling under a time-of-use tariff; chiller staging with condenser water reset; chilled water supply temperature reset with a wet cooling coil; and chilled water storage under time-of-use prices and a demand charge. The plants were modelled with the Modelica Buildings library, and in three of the cases a soft actor-critic agent was trained directly on the exported functional mock-up unit (FMU). Each agent was compared with tuned rule-based controllers and, where practicable, with a near-optimal reference from hourly exhaustive search or dynamic programming with perfect foresight. In the two cases without storage the agent reached the near-optimal reference, but the margin above a well-chosen rule was small: 1.3% for condenser water reset and 0.18% for chilled water reset. In the latter case two common reset rules consumed more energy than a fixed set point. With a chilled water storage tank the margin was considerably larger. The agent outperformed both rules (-7.8%, compared with -3.9% and -6.3%) and adopted midday charging during the solar off-peak period, yet it obtained only about 45% of the saving achieved by the perfect-foresight schedule and did little to limit the demand peak. In the pre-cooling case a hand-tuned rule remained superior (-12.0% compared with -5.6%). These results suggest that studies of DRL for HVAC control should state the margin between the best rule and a near-optimal reference. Some practical difficulties encountered in training on co-simulation FMUs are also reported.

cs.LG · cs.AI · eess.SY

版本更新 · v2arXiv:2610.18352

Let the Mind Conceive, Let the Hand Achieve: Coupling Grounded Action and Progress Evaluation for Long-Horizon Manipulation

Youpeng Wen, Gang Wang, Tianyu Huang, Weiwei Zhang, Zhijian Huang, Wei Chen, Xieyuanli Chen, Yunhui Liu

原文摘要

A robot following a multi-step instruction needs to act on the right objects in the right order, even as the scene changes. It must also maintain awareness of task progress and determine when to advance. To address these challenges, we introduce Orchestra, an agent harness that separates task planning and execution management from action generation. Orchestra decomposes a complete multimodal instruction into ordered subtasks, tracks completed steps, and updates the locations of relevant objects and destinations during execution. It maintains a shared context containing the active subtask, spatial grounding, and recent observations. An action policy uses this context to generate actions, while a separately trained progress evaluator estimates when to advance to the next subtask. To evaluate whether robots trained on atomic instructions and short trajectories can execute longer subtask sequences, we introduce a benchmark named RMBench-Compose. Orchestra achieves a mean success rate of $68.22\%$ across nine RMBench tasks, outperforming Mem-0 by $26.22$ percentage points. On RMBench-Compose, it achieves an Average Success Length (ASL) of $7.35$ subtasks, $2.46$ times that of HarnessVLA. Across four real-world multi-step tasks, it improves mean ASL by $25.0\%$ over Orchestra using the base $π_{0.5}$ policy. These results show that Orchestra helps robots carry out longer sequences of subtasks while preserving the requested subtask order and acting on the specified objects. Our code and project page are available at https://youpengwen.github.io/Orchestra

cs.RO · cs.AI

版本更新 · v2arXiv:2605.21752

PEARL: Unbiased Percentile Estimation via Contrastive Learning for Industrial-Scale Livestream Recommendation

Blake Gella, Wei Wu, Yuhao Yin, Zexi Huang, Zikai Wang, Emily Liu, Junlin Zhang, Wentao Guo 等 9 位作者

原文摘要

Recommender systems trained on user interaction data are susceptible to behavioral intensity imbalance--a systematic distortion arising from heterogeneous engagement patterns across users. This imbalance skews feedback signals such that observed interactions no longer faithfully reflect true preferences, causing models to disproportionately amplify signals from highly active users while underrepresenting others, which ultimately degrades recommendation quality and robustness at scale. To address this issue, we propose a nonparametric contrastive percentile approximation framework, PEARL, that models relative preference signals instead of absolute engagement magnitudes. Building upon relative advantage debiasing, PEARL leverages real contrastive interaction samples to approximate percentile relationships directly, without relying on auxiliary distribution estimation models. We provide theoretical justification demonstrating that such pairwise comparisons yield unbiased estimates of percentile-based preference signals. For broader applicability, we introduce a prediction-based bootstrapping mechanism for percentile smoothing to handle sparse and discrete feedback, alongside a generalized value-weighted formulation and a co-training strategy to enhance both modeling flexibility and representation learning. Extensive offline experiments demonstrate that PEARL effectively mitigates behavioral bias and consistently improves recommendation performance across multiple ranking targets. Deployed in a production livestream platform with a combined user base of billions, online A/B testing confirms substantial real-world gains: +2.10% Watch Duration, +0.80% Consumption Amount, +1.49% Interaction Rate, and -6.91% Report Rate.

cs.LG · cs.AI