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

+1680新增论文
+1333其中 AI 相关
+187已有论文修订
+59新增 AI 导读

arXiv 最近完成扫描:2026-11-18 04:28 · 全站最近入库:2026-11-17 23:35各信源进度 →

统计口径与数据新鲜度

统计 2026-11-17 06:00 — 2026-11-18 06:00(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:2026-11-18 04:28最近发现新增版本:2026-11-17 23:35库内论文最新更新时间:2026-11-17 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。

25,786 已收录论文21,491 AI 相关7,725 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 30 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-18 04:27(北京时间)
新论文 · v1arXiv:2611.21165

A Matter of Stages: Energy-Latency Trade-offs in Multimodal LLM Serving

Yifan Sun, Jiantong Jiang, Zhexiang Zhang, Xu Bai, Feng Liu, Rajkumar Buyya, Adel N. Toosi

原文摘要

Online serving of multimodal large language models requires deciding whether to separate visual encoding (E), prefill (P), and decode (D) across GPUs and how to configure their frequencies. EPD disaggregation enables independent stage control, but it also changes communication and queueing behavior. Lowering one stage's frequency can increase queueing and prolong execution across the EPD pipeline. Consequently, reducing power at one stage does not necessarily reduce total GPU energy per request. In this paper, we systematically characterize fused and EPD-disaggregated serving across four multimodal models under a fixed three-GPU budget, followed by single-stage and uniform downclocking experiments across 18 workload combinations. We find that disaggregation reduces GPU energy per successful request by 13.3-37.0% at nearly identical output-token throughput, but increases mean TTFT. Downclocking exhibits distinct stage sensitivities, and aggressive frequency reductions can offset power savings through queueing and longer execution. Across the evaluated workloads, uniform downclocking to 80% reduces GPU energy by 11.7-18.6% with modest mean-latency changes. Guided by these findings, we develop a budgeted frequency exploration strategy to select energy-efficient configurations under latency constraints. Evaluation shows that four probes can identify the lowest-energy feasible configuration among 25 measured candidates across all 288 evaluated positive-margin SLO scenarios. These results provide practical guidance for energy-efficient multimodal serving.

cs.DC

新论文 · v1arXiv:2611.21072

Sage S2.1 Flash Technical Report

Kun Cai, Haodong Chen, Guangran Cheng, Qingxiu Cheng, Xu Dai, Guanting Dong, Qiming Ge, Lixin Gu 等 206 位作者

原文摘要

We introduce Sage S1.5 Flash, a 320-billion-parameter general-purpose mixture-of-experts model designed to reason over scientific evidence and act in interactive environments. Sage Sci-Perception uses dedicated encoders for molecular graphs and biological sequences and represents spectra as ordered patches of measurements. Modality-specific projectors map these representations into the input space of a shared backbone, which processes them alongside text and images. To develop the model's agent capabilities, we construct verifiable environments for tool use, search, and coding, as well as specialized workflows involving machine learning experiments and scientific software. Task completion is checked against execution results, changes in environment state, or submitted artifacts. We evaluate Sage S1.5 Flash on 33 benchmarks across seven domains. In the primary comparison against five baselines, it ranks first on OfficeQA Pro (72.2), MLE-bench Lite (90.3 HumanRank), Harvey LAB, and SpreadsheetBench~2, and second on OSWorld-Verified (85.2), WebArena-Verified (67.6), and DeepResearch Bench~II (54.2). These results show particular strengths in using documentary evidence, producing verifiable workbooks, and carrying out machine learning tasks, alongside competitive performance in desktop and browser workflows that require the model to act on visual feedback.

cs.AI

版本更新 · v2arXiv:2610.41291

In-Model Context Compression with Agentic Soft Tokens for Long-Horizon LLM Agents

Xuning He, Kai Fan, Runyu Zhang, Binghuan Wang, Yuxiang Zhang, Xue Jiang, Yihong Dong

原文摘要

Long-horizon LLM agents play an important role in solving complex and sustained tasks, with ever-growing interaction contexts. However, these accumulated contexts will raise substantial computational cost and obscure key decision-relevant evidence. Existing methods typically train external compressors or context-management agents to rewrite context as shorter hard-token sequences, but such external selection may not align with the LLM's evolving representation and is necessarily lossy. In this paper, we introduce Agentic Soft-Token Compression (ASTC), a training-free compression method using soft tokens within LLM agent for long-horizon tasks. Specifically, ASTC uses the model's native recurrent dynamics to determine the already absorbed information, leveraging the commonly-used hybrid Gated Delta Network (GDN)-attention. It then carries this information forward as agentic soft tokens while retaining exact values as hard tokens. The resulting compact sequence reduces later-layer computation and yields reusable KV caches, extending the savings across the agent's trajectory. Extensive experiments on long-horizon tasks show that ASTC matches or exceeds the uncompressed agent's accuracy with up to 59.84% fewer end-to-end Attention Pairs, while trained compressors require 12.3 to 194.9 times more Attention Pairs per successful task. These findings highlight agentic soft tokens as a model-native way to carry absorbed context forward as the task representation evolves.

cs.CL · cs.AI · cs.LG · cs.NE

新论文 · v1arXiv:2611.19684

Early Readability of Trajectory Quality in Diffusion Driving Planners

Xiang Zhang, Ziyang Leng, Yuxin Bao, Zhihao Zhao, Bolei Zhou, Xugui Zhou

原文摘要

Diffusion-based driving planners can generate multiple trajectories for the same scene, but these candidates vary in quality. Can their internal states reveal which candidate will be better before denoising is complete? We systematically study this question in three pretrained planners: Alpamayo, ReCogDrive-IL, and HDP, using linear probes to predict final trajectory quality from intermediate states and evaluating candidate ranking and selection on held-out scenes. Across all three planners, we observe that states from the first denoiser evaluation already support informative selection, while further denoising does not consistently improve selection quality. Further analyses show that readout depends on the representation and reader, and that selection can coexist with limited overall ranking accuracy. These findings motivate an early-selection method that selects after the first denoiser evaluation and completes only the chosen candidate. This reduces computation relative to fully denoising every candidate and improves open-loop trajectory quality over random selection across all three planners. In particular, it reduces Alpamayo's average displacement error by 17.8% and 13.4% in one- and four-camera configurations, respectively. Closed-loop evaluations also show higher mean driving scores and progress.

cs.RO

新论文 · v1arXiv:2611.18417

CoPre-MoE: Consolidating and Preserving MoE for Multimodal Continual Instruction Tuning

Tao Hu, Ling-Xiang Zhang, Da-Wei Zhou

原文摘要

Multimodal continual instruction tuning requires multimodal large language models to sequentially acquire new capabilities without catastrophic forgetting. Mixture-of-experts (MoE) architectures offer a promising way to mitigate task interference through expert specialization and combination. However, while existing MoE-based methods avoid overwriting shared parameters by allocating separate experts for each task, this progressive expansion incurs severe cross-task redundancy and linear parameter growth. Specifically, experts across tasks often learn overlapping functions alongside task-specific capabilities. For instance, ScienceQA and TextVQA both locate relevant image regions. To curb parameter redundancy without inducing forgetting, we propose Consolidating and Preserving MoE (CoPre-MoE). For each incoming task, CoPre-MoE first learns task-specific MoE to capture specialized capabilities. To eliminate redundancy and reduce parameter growth across tasks, we design a shared-and-private expert architecture. Specifically, we first evaluate functional overlap to selectively consolidate overlapping experts across tasks into a fixed-capacity generic expert pool. We then factorize the remaining task-specific experts into shared projection bases and compact task-specific cores within a private expert pool. To mitigate forgetting, we construct the router by pairing each expert with learnable retrieval keys, where historical keys remain frozen during continual updates while newly added keys are regularized with an orthogonality penalty. At inference, input-dependent routing dynamically combines both pools without test-time task labels. Extensive experiments demonstrate that CoPre-MoE achieves superior performance while substantially reducing parameter overhead.

cs.CV · cs.AI · cs.LG

新论文 · v1arXiv:2611.17287

面向单边可验证奖励的仅负样本策略优化

Negative-Only Policy Optimization for One-Sided Verifiable Rewards

Jiacheng Xu, Shuo He, Fuxiang Zhang, Chaojie Wang, Bo An

AI 速读 基于摘要

针对只能可靠证伪、无法可靠验证正确的单边可验证奖励场景,提出仅利用可信负样本更新策略的 NOPO 方法。

cs.LG · cs.CL

查看原文摘要

Reinforcement learning with verifiable rewards (RLVR) typically relies on a verifier that can reliably judge whether sampled solutions are correct. However, in many realistic settings, complete verification is generally unavailable, while cheap checks can only falsify failures. We formalize this regime as one-sided verifiability (OSV): verified negatives are reliable counterexamples, but verified positives are merely non-falsified candidates and may contain many false positives. This asymmetry makes standard RLVR unreliable, because reinforcing verified positives can directly reward shortcuts that pass partial checks. To address this issue, we propose Negative-Only Policy Optimization (NOPO), which treats verified positives as unlabeled and updates the policy only by suppressing trusted verified negatives. Specifically, NOPO combines a reference-relative negative suppression loss with a group-level adapter that scales each update by the proportion of non-falsified responses. Across unit-test, constraint-check, and LLM self-verifiers, NOPO improves pass@k over baselines on code, logic, and math benchmarks when oracle rewards are unavailable. These gains persist under inference-time scaling and cross-benchmark transfer.

版本更新 · v2arXiv:2610.38520

What Hallucinations Have We Measured? Semantic Illusion, Confabulation, and Expressed Uncertainty

Hazel H. Kim, Jakub Binkowski, Ekaterina Fadeeva, Juan David Guerra, Yijie Hao, Zheng Yi Ho, Baixiang Huang, Buyun Liang 等 26 位作者

原文摘要

Hallucination remains one of the most critical challenges in deploying large lan- guage models (LLMs) reliably and safely. Yet the term is used inconsistently across the research community. We argue for a strict definition of hallucinations in LLMs. First, hallucinations should be confident incorrectness, kept separate from expressed uncertainty (UNK). UNK reveals that the model may be wrong, whereas a hallucination does not reveal it and is therefore more dangerous in prac- tice. Conflating the two obscures their distinct behaviors and makes them hard to study. Second, hallucinations should be split by a simple re-elicitation test: does any semantically equivalent prompt recover the correct response? This test probes whether the relevant knowledge is accessible. If none does, the knowledge is likely absent or inaccessible, and the error type is a confabulation (CF) that likely requires external knowledge or retraining; if some does, the knowledge is accessible from the model but applied inconsistently, and the error type is a semantic illusion (SI) that simple prompt or decoding changes can mitigate3. Re-evaluating 14 recent hallucination methods under this taxonomy yields three findings. (1) Mitigation methods are far more effective for SI than for CF. (2) Detection methods flag UNK more often than hallucinated responses (i.e., SI and CF). (3) Latent signals are more informative than output-level signals for CF. An audit of recent hallucination papers shows that most conflate these categories. We recommend reporting CF, SI, and UNK separately, and introduce a practical labeling protocol that needs only a modest number of semantically equivalent prompts.

cs.CL

新论文 · v1arXiv:2611.13907

EVAThink: Evidence Sufficiency-Guided Training for Adaptive Thinking in Video Understanding

Changti Wu, Ruizhe Fan, Bin Wang, Lei Zhang, Yuxiang Zhang

原文摘要

Under a fixed frame budget, video understanding with vision-language models (VLMs) depends on whether the sampled frames provide sufficient evidence for the question. Sufficient frame evidence supports direct answering, whereas less informative observations may benefit from reasoning that connects the available cues. We introduce EVAThink, which enhances video understanding through evidence sufficiency-guided training for adaptive thinking. First, Recoverable Evidence Gap Mining constructs evidence-dense and evidence-sparse video views that share the same question and answer under an identical frame budget. It retains pairs where sparse views yield lower direct-answer accuracy than dense views but benefit from explicit reasoning, providing supervision for response-mode selection. Second, Evidence-Guided Two-Stage Reinforcement Learning uses stage-specific rewards to guide training with evidence-based mode preferences and then refine response efficiency, while first-token advantage regularization mitigates mode collapse. Trained on only 6.7K curated samples, EVAThink consistently improves upon the base model across video understanding benchmarks under both 32- and 64-frame budgets, while outperforming strong baselines on average. The code, model, and datasets are available at \href{https://github.com/ChangtiWu/EVAThink}{EVAThink}.

cs.CV

新论文 · v1arXiv:2611.12572

Text2NGAC: a multi-LLM pipeline for NGAC policy extraction with mutation-based verification

Ayesha Habib, Huixiang Zhang, Dan Zhu, Waiz Haider

原文摘要

Access-control policies are written in natural language but enforced as formal models, and translating them into NIST Next Generation Access Control (NGAC) graphs remains unsolved: prior systems cover only subsets of the NGAC model, and none verify the graphs they extract. We present Text2NGAC, a pipeline of cooperating large language models that disambiguates each sentence before extraction, escalates uncertain sentences to a multi-model ensemble, and checks the resulting graph with a five-strategy verifier whose own effectiveness is measured by controlled mutation testing. To our knowledge it is the first system to extract all four NGAC relationship types - associations, prohibitions, event-response obligations, and assignments - from real-world policy documents. Three results anchor the evaluation. The verifier detects 99.7% of 918 injected mutants and every structural fail-open fault, and the fault classes it misses are characterized and weighted by security consequence rather than left unexamined. The extracted graphs are enforceable: one corpus's graph loads unmodified into the NIST reference implementation, which reproduces all 55 access decisions the graph entails and fires an extracted obligation end-to-end. Extraction reaches 97.3-99.6% exact-tuple-match accuracy on the 714 scored sentences of four real corpora, far above three reproduced baselines on matched metrics, and paraphrase probes with semantic re-scoring bound the pipeline's sensitivity to rewording. Reported accuracy is wording- and benchmark-specific, adversarial input is outside the verifier's guarantees, and the intended deployment is extraction plus structural verification feeding expert review. A complete two-tier reproducibility artifact accompanies the paper.

cs.CR

版本更新 · v2arXiv:2610.12468

DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training

Junyan Li, Ruizhi Li, Yu Liu, Xiangshuo Liu, Mingchao Sun, Hongyu Pan, Mu Xu, Lue Fan 等 9 位作者

原文摘要

We present DreamTrue, a multi-view, cross-embodiment robot world model for action-faithful and physically plausible video prediction. Training such a model on existing robot datasets faces two obstacles: imprecise calibration can impair action following, while limited coverage of unsuccessful interactions can bias predictions toward successful outcomes. To improve action following across embodiments, we render action trajectories into image-space conditions and introduce offline geometric calibration to align these conditions with the target videos. To broaden interaction coverage, we introduce counterfactual post-training, modifying recorded action trajectories and generating future videos under a wider range of actions and contact configurations. To provide feedback on these predictions without paired ground-truth futures, we construct a human-annotated video dataset covering robot, object, and interaction defects and use it to train an embodied video reward model. Its scores guide reinforcement-learning post-training toward more physically plausible interaction outcomes. On AgiBot, DreamTrue attains state-of-the-art action following, while reducing the human-assessed interaction defect rate from from 48.12% to 6.25%. Notably, our model ranks first in the world model track of the AgiBot World Challenge 2026. The project page can be found at https://brave-eai.github.io/DreamTrue.

cs.RO · cs.CV

版本更新 · v3arXiv:2602.05243

CORP: Closed-Form One-shot Representation-Preserving Structured Pruning

Boxiang Zhang, Baijian Yang

原文摘要

Transformers achieve strong accuracy but incur high compute and memory cost. Structured pruning reduces inference cost, but most methods rely on retraining or multi-stage optimization, which limits post-training deployment. We propose CORP, a closed-form one-shot structured pruning method that removes MLP dimensions and attention substructures using only unlabeled calibration data without gradients or fine-tuning. CORP formulates structured pruning as a representation recovery problem. It models removed components as affine functions of retained components and derives closed-form ridge regression solutions that fold compensation into model weights. This minimizes a layer-local affine/logit reconstruction objective under the calibration distribution. Experiments on ImageNet with DeiT reveal strong redundancy in both MLP and attention representations. With CORP, models retain high accuracy under aggressive sparsity. On DeiT-Huge, CORP achieves 83.24% Top-1 accuracy after pruning 50% of both MLP and attention structures.

cs.LG · cs.CV

新论文 · v1arXiv:2611.10542

When Experience Becomes a Backdoor: Poisoning Self-Evolving LLM Agents via Failure-Aware Task Optimization

Minhua Lin, Zongyu Wu, Hanqing Lu, Xianfeng Tang, Xiang Zhang, Qi He, Suhang Wang

原文摘要

Large language model (LLM) agents increasingly improve themselves through self-evolution, distilling execution experience into persistent, reusable skills. However, this persistence creates a new attack surface: malicious rules implanted into skills can compromise subsequent agent behavior. Existing skill backdoor attacks typically assume direct manipulation of the victim's skills, which is impractical when skills are constructed internally by a trusted evolver. In this paper, we study a novel problem of task-level backdoor attacks, where the attacker can only inject poisoned evolution tasks, inducing the agent to implant backdoors into its own skills. Since the attacker controls neither skill induction nor downstream execution, a natural approach is to iteratively optimize poisoned tasks based on attack success. However, our preliminary analysis reveals two key challenges: (i) failures at different stages of skill evolution and downstream execution can produce indistinguishable outcomes, providing little guidance for targeted poisoned-task updates; and (ii) unsuccessful updates may undermine earlier progress, while simply discarding them may lead to repeated ineffective updates. To address these challenges, we propose BadEvolve, a failure-aware task-poisoning framework against self-evolving LLM agents. BadEvolve diagnoses attack failures across four stages, i.e., Induction, Activation, Following, and Effect, and directs updates toward the failing stage. It further employs rollback to preserve earlier progress and failure memory to avoid repeating ineffective updates. Extensive experiments across diverse self-evolution frameworks, benchmarks, and backbone LLMs demonstrate the effectiveness of BadEvolve. Our code is available at https://github.com/ventr1c/BadEvolve.

cs.CR · cs.AI

版本更新 · v2arXiv:2510.03224

Test-Time Defense Against Adversarial Attacks via Stochastic Resonance of Latent Ensembles

Dong Lao, Yuxiang Zhang, Haniyeh Ehsani Oskouie, Yangchao Wu, Alex Wong, Stefano Soatto

原文摘要

We propose a test-time defense mechanism against adversarial attacks: imperceptible image perturbations that significantly alter the predictions of a model. Unlike existing methods that rely on feature filtering or smoothing, which can lead to information loss, we propose to "combat noise with noise" by leveraging stochastic resonance to enhance robustness while minimizing information loss. Our approach introduces small translational perturbations to the input image, aligns the transformed feature embeddings, and aggregates them before mapping back to the original reference image. This can be expressed in a closed-form formula, which can be deployed on diverse existing network architectures without introducing additional network modules or fine-tuning for specific attack types. The resulting method is entirely training-free, architecture-agnostic, and attack-agnostic. Empirical results show state-of-the-art robustness on image classification and, for the first time, establish a generic test-time defense for dense prediction tasks, including stereo matching and optical flow, highlighting the method's versatility and practicality. Specifically, relative to clean (unperturbed) performance, our method recovers up to 68.1% of the accuracy loss on image classification, 71.9% on stereo matching, and 29.2% on optical flow under various types of adversarial attacks.

cs.CV · cs.AI · cs.LG

版本更新 · v3arXiv:2606.01869

WorldCoder-Bench: Benchmarking Physically Grounded 3D World Synthesis

Shuo Lu, Yinuo Xu, Kecheng Yu, Siru Jiang, Yongcan Yu, Yubin Wang, Haitao Yang, Yuxiang Zhang 等 11 位作者

原文摘要

Large language models (LLMs) are increasingly asked not only to write static interfaces, but to construct executable interactive worlds from natural language. Browser-native 3D, commonly built with Three.js, is a natural next frontier: generated programs must integrate assets, obey spatial and physical constraints, and keep user-facing controls synchronized with hidden runtime state. Existing web-generation benchmarks and evaluators, however, largely observe only pixels or DOM nodes, while the mechanics of a Three.js world unfold inside an opaque <canvas>. We introduce WorldCoder-Bench, a benchmark for autonomous, physically grounded 3D world synthesis. WorldCoder-Bench contains 2,026 expert-curated tasks across Simulation, Rendering, and Application scenarios, with optional .glb assets and hidden behavioral contracts. We further propose StateProbe, an execution-based protocol that probes generated programs in a controlled browser and verifies hidden, mutation-hardened contracts over runtime states and transitions. Beyond verification coverage, we report Return on Automation and Time Efficiency Multiplier to estimate quality-adjusted cost and time utility. The best verification coverage is 27.8% across nine models on WorldCoder-Benchcore and 19.9% across five on WorldCoder-Benchrob, with failures dominated by state-schema drift and broken interaction chains rather than missing scene elements. Utility metrics further show that cheap or fast models can still provide substantial value on easier domains. WorldCoder-Bench is available at https://github.com/shuolucs/WorldCoder-Bench.

cs.AI

版本更新 · v4arXiv:2607.11007

TabPFN beyond Tabular Data: Calibration and Accuracy on Multimodal Embeddings

Jingxiang Zhang, Lujia Zhong, Zijie Zhu, Shuo Huang, Yuang Xu

原文摘要

Few-shot multimodal classification commonly attaches a lightweight head, such as $k$-nearest neighbors, logistic regression, or a linear SVM, to a frozen pretrained encoder. Although computationally efficient, these heads can produce poorly calibrated confidence scores. We ask whether TabPFN can provide reliable confidence estimates on multimodal embeddings without sacrificing predictive accuracy, and under what conditions. We systematically evaluate TabPFN as a zero-gradient head for frozen image, text, and audio encoders. Across 22{,}820 evaluation episodes spanning 14 datasets, 11 encoders, and three modalities, TabPFN achieves the best mean rank among nine classification heads on both negative log-likelihood (NLL) and expected calibration error (ECE). At a representative setting, it reduces NLL by 48--62\% and ECE by 2.1--5.3$\times$ relative to the average of eight baselines while matching or exceeding their average accuracy. This calibration benefit transfers broadly, whereas the accuracy advantage is conditional: it concentrates at moderate-to-high shot counts and low-to-moderate feature dimensions ($k \ge 50$, $d \le 32$), and diminishes when labeled data are scarce, feature dimensions are high, or competing methods approach ceiling accuracy. After backbone adaptation, replacing the trained linear head with TabPFN improves calibration while preserving competitive accuracy, showing that representation adaptation and reliable head choice are complementary. Together, these results identify when TabPFN can serve as a training-free head for calibration-sensitive multimodal classification. To support transparency and reproducibility, we publicly release the source code, experiment configurations, and evaluation scripts in our GitHub repository: https://github.com/Jingxiang-Zhang/tabpfn-multimodal-embeddings.

cs.LG

版本更新 · v3arXiv:2602.00359

Position: Agentic Evolution is the Path to Evolving LLMs

Minhua Lin, Hanqing Lu, Zhan Shi, Bing He, Rui Mao, Zhiwei Zhang, Zongyu Wu, Xianfeng Tang 等 15 位作者

原文摘要

As Large Language Models (LLMs) move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world. Our code is publicly available at https://github.com/A-EVO-Lab/a-evolve.

cs.AI

版本更新 · v2arXiv:2605.30621

Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents

Minhua Lin, Juncheng Wu, Zijun Wang, Zhan Shi, Yisi Sang, Bing He, Zewen Liu, Tianxin Wei 等 17 位作者

原文摘要

LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.

cs.AI

版本更新 · v3arXiv:2606.00386

αDepth: Learning Scene-Level Soft Boundary Decomposition for Stereo Conversion

Xiang Zhang, Yang Zhang, Lukas Mehl, Karlis Martins Briedis, Markus Gross, Christopher Schroers

原文摘要

Accurately modeling soft boundaries, e.g., hair and defocus blur, is a fundamental challenge in stereo conversion due to the ambiguous blending of foreground and background. Existing depth models primarily predict single-layer depth, leading to ambiguity in depth correspondence at soft boundaries. While matting techniques can capture opacity for layered modeling, they often struggle in complex scenes with multiple targets and usually require user intervention. This paper introduces αDepth, a layered representation that decomposes soft boundaries for high-fidelity stereo conversion. Specifically, we first resolve mixed color and depth ambiguity by estimating layered color and depth values at soft boundaries. Considering complex multi-target scenes, we design a Circular Alpha Representation (CAR) that shifts the paradigm from global target extraction to local boundary decomposition. Unlike prior matting methods restricted to a single foreground/background, CAR enables efficient scene-level inference without manual guidance. Extensive evaluations demonstrate that αDepth achieves state-of-the-art performance in stereo conversion, eliminating background bleeding and structural distortions at soft boundaries.

cs.CV

版本更新 · v2arXiv:2611.01560

让机器人使用智能体走向大众:一个配开放模型的统一测试框架

Democratizing Robot-Use Agents: A Unified Harness with Open Models

Jiarui Yuan, Yuxin Zuo, Xiaoyu Huang, Wenbo Zhang, Kwanyoung Park, Junlin Yang, Tianxiang Zhang, Ruoqu Chen 等 28 位作者

AI 速读 基于摘要

作者提出统一测试框架 uni-robot-agent,使开放权重模型可开展机器人使用智能体研究并支持后续训练,报告开箱即用的 Qwen3.8-Flash-Next 在 RoboDojo 上取得 31.39 分。

cs.RO

查看原文摘要

Robot-use Agents (RUAs) have shown promising zero-shot manipulation capabilities, but reliance on frontier, proprietary models limits research into further training. To democratize RUA research, we introduce uni-robot-agent, a unified harness for competitive open-weight robot use that supports further training. Without privileged inputs, an out-of-the-box Qwen3.8-Flash-Next achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. Our analysis across model generations, families, and sizes links effective robot use to four general capabilities, suggesting that future RUA training should develop them jointly.

版本更新 · v2arXiv:2610.23344

Language Models Learn to Improve Language Models

Minhua Lin, Zijun Wang, Zewen Liu, Zhaotian Weng, Bing He, Zhan Shi, Benoit Dumoulin, Xiang Zhang 等 10 位作者

原文摘要

Large language models (LLMs) are increasingly used to automate AI research, including language-model (LM) development, offering a potential path toward recursive self-improvement. However, turning experimental feedback into a learned ability to propose effective improvements across models faces two challenges. First, training such an ability with real feedback is expensive: reinforcement learning (RL) requires evaluating tens of thousands of proposals, each potentially requiring a full model-training run. Even at ten GPU-hours per run, reward collection alone would consume hundreds of thousands of GPU-hours, motivating a cheaper surrogate. Second, surrogate feedback can become unreliable as policy optimization favors overestimated proposals when the proposal distribution shifts. To address these challenges, we propose LM2LM, which trains LLMs with RL to generate effective LM improvement proposals. To provide scalable feedback, we use a research world model (RWM), a frozen LLM that predicts proposal gains from real experimental records provided in context, as the reward model for RL. To maintain feedback reliability, we periodically execute current-policy proposals ranked highly by the RWM and add their measured outcomes to its context. Across eight autoregressive and diffusion LM development environments, real-execution evaluations on held-out tasks show that LM2LM enables Qwen3.8-27B to outperform GPT-6-sol, even when GPT-6-sol is given 300 real experimental records in context, while requiring 33% lower data-generation cost. LM2LM also generalizes to a model family unseen during training.

cs.CL · cs.AI

新论文 · v1arXiv:2611.01529

AR-RRG: Anatomy-Guided Retrieval-Driven Radiology Report Generation

Xiang Zhang, Shijie Song, Wanqing Zhao, Wei Zhou, Rui Cao, Dekui Wang, Qiyao Hu, Xianlin Peng 等 11 位作者

原文摘要

Radiology report generation (RRG) streamlines clinical workflows and alleviates the burden of radiologists' documentation. Standard vision-language models capture global visual cues and lack fine-grained alignment between anatomical regions and diagnostic text. Most retrieval-augmented methods retrieve complete historical reports, introducing redundant noise and severe vision-text mismatches that trigger clinically risky hallucinations and missed lesions. In this paper, we introduce AR-RRG, an anatomy-guided retrieval-driven method for the generation of radiology reports. AR-RRG consists of two core modules: the anatomy-aware self-training lesion detection network (ASLDNet) and the retrieval-driven radiology report generation (RRRG). ASLDNet leverages anatomy-aware spatial verification (ASV) and class-adaptive dynamic thresholding (CDT) via self-training, delivering accurate lesion localization to support subsequent generation. As the generative branch of AR-RRG, RRRG eliminates clinical hallucinations through dual-granularity retrieval (DR) and complements missing lesions with a neighborhood consensus (NC) mechanism. A parallel continuous visual perception (PCVP) stream converts radiographs into visual tokens through fixed medical encoding and resampling. Multi-scale text and visual features are fused, followed by LoRA-efficient large language model tuning to generate clinically consistent, visually grounded radiology reports. Extensive evaluations on MS-CXR, MIMIC-CXR, and IU X-Ray demonstrate that AR-RRG consistently outperforms state-of-the-art methods across mainstream metrics. Our code will be made publicly available.

cs.CV

新论文 · v1arXiv:2610.36840

Trace-WAM: Learning to Act via 3D Motion Prediction for World Action Models

Xu Hu, Changjie Wu, Linzhe Shi, Hang Zhang, Chen Yang, Lingjun Zhang, Guowen Zhang, Mu Xu 等 11 位作者

原文摘要

World Action Models (WAMs) learn representations for robot control by coupling future visual prediction with action generation. Successful manipulation depends primarily on how the robot and task-relevant objects evolve, whereas future-video prediction must also reconstruct the appearance and background of the entire scene. We therefore propose Trace-WAM, which augments video-action learning with future 3D motion prediction as an auxiliary objective. This objective supervises temporal point correspondence together with image-plane displacement and depth change, directly describing the interaction dynamics and spatial changes involved in manipulation. To this end, we adopt a three-branch Mixture-of-Transformer (Tri-MoT) architecture to jointly learn video prediction, motion prediction, and action generation. The motion expert predicts future trajectories from current visual features, allowing motion supervision to shape the representations used for action generation. Because the action expert does not consume predicted trajectories, inference retains the current-observation policy path: a single visual forward pass followed by action denoising. Experiments on RoboCasa-GR1, LIBERO, and LIBERO-Plus demonstrate strong in-distribution performance and out-of-distribution generalization, with particularly clear gains over Fast-WAM under limited demonstrations and distribution shifts.

cs.RO

版本更新 · v2arXiv:2605.17912

WorldArena 2.0: Extending Embodied World Model Benchmarking on Modality, Functionality and Platform

Zhuohang Li, Yiding Ma, Yu Shang, Lei Jin, Shengyuan Wang, Jianjie Fang, Zhaolu Wang, Xvyuan Liu 等 29 位作者

原文摘要

World models have emerged as a central paradigm for embodied intelligence, enabling agents to predict action-conditioned future and reason about environmental dynamics. However, existing embodied world model benchmarks are still largely confined to vision-only prediction, offline embodied applications, and simulator-based evaluation, making them insufficient for assessing increasingly comprehensive world models. In this work, we introduce WorldArena 2.0, an expanded benchmark that systematically broadens embodied world model evaluation along three dimensions: modality, functionality, and platform. Along the modality dimension, WorldArena 2.0 extends evaluation from vision-only to visuotactile modalities, enabling assessment of multimodal perception and prediction. Along the functionality dimension, it extends beyond policy evaluation and planning to assess world models as interactive RL environments for policy optimization. Along the platform dimension, it moves beyond simulator-only evaluation to a diverse suite of simulated and real-world robotic settings across multiple embodiments. Under a standardized protocol, WorldArena 2.0 comprehensively evaluates perceptual quality, interactive utility, and cross-platform performance, providing a comprehensive testbed for tracking progress toward embodied world models. The project page including codes and demos is available at: https://world-arena.ai/.

cs.RO · cs.CV

版本更新 · v4arXiv:2609.31494Google Research

面向可证明隐私的联邦数据学习

Toward provably private learning from federated data

Katharine Daly, Yu Xiao, Zachary Garrett, Brett McLarnon, Jianpeng Hou, Arun Ganesh, Yanxiang Zhang, Noriyuki Takahashi 等 12 位作者

AI 速读 基于摘要

提出基于可信执行环境(TEE)的下一代联邦学习系统,首次提供可外部验证的中心化差分隐私(DP)保证,并改善隐私-效用权衡。

cs.CR

查看原文摘要

Federated Learning (FL) allows devices with private data to collaborate in training a shared model. We present a next-generation FL system based on Trusted Execution Environments (TEEs) that addresses operational challenges associated with earlier systems and provides externally verifiable central Differential Privacy (DP) guarantees for the first time while offering a better privacy-utility tradeoff. In our system, devices upload data encrypted with keys managed by a TEE-hosted Key Management Service (KMS). The uploaded data is cryptographically tied to a policy limiting the set of Python programs that may later process the data in server-side TEEs. External parties may inspect public transparency logs to observe the set of workloads allowed by these policies. Our experimental results show that the new system improves device coverage and favorably shifts privacy-utility curves by enabling collected data to be integrated into the server-side workload at a schedule that optimizes DP guarantees and is unaffected by device availability. Our new system has been productionized, enabling models for the Android Keyboard (Gboard) to be trained faster and achieve better accuracy under smaller, now externally verifiable privacy budgets in comparison to models trained using the prior system.

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.139ACL Anthology · ACL 2026

CriticLean: Critic-Guided Reinforcement Learning for Mathematical Formalization

Zhongyuan Peng, Yifan Yao, Kaijing Ma, Shuyue Guo, Yizhe Li, Yichi Zhang, Chenchen Zhang, Yifan Zhang Princeton University 等 19 位作者

原文摘要

Translating natural language mathematical statements into formal, executable code is a fundamental challenge in automated theorem proving. While prior work has focused on generation and compilation success, little attention has been paid to the critic phase—the evaluation of whether generated formalizations truly capture the semantic intent of the original problem. In this paper, we introduce CriticLean, a novel critic-guided reinforcement learning framework that elevates the role of the critic from a passive validator to an active learning component. Specifically, first, we propose the CriticLeanGPT, trained via supervised fine-tuning and reinforcement learning, to rigorously assess the semantic fidelity of Lean 4 formalizations. Then, we introduce CriticLeanBench, a benchmark designed to measure models’ ability to distinguish semantically correct from incorrect formalizations, and demonstrate that our trained CriticLeanGPT models can significantly outperform strong open- and closed-source baselines. Building on the CriticLean framework, we construct FineLeanCorpus, a dataset comprising over 509K problems that exhibits rich domain diversity, broad difficulty coverage, and high correctness based on human evaluation.Overall, our findings highlight that optimizing the critic phase is essential for producing reliable formalizations and we hope our CriticLean will provide valuable insights for future advances in formal mathematical reasoning.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.1477ACL Anthology · ACL 2026

ReviewGrounder: Improving Review Substantiveness with Rubric-Guided, Tool-Integrated Agents

Zhuofeng Li, Yi Lu University of Waterloo and Texas A&M University - College Station, Dongfu Jiang, Haoxiang Zhang, Yuyang Bai, Chuan Li Lambda, Yu Wang University of Oregon and Vanderbilt University, Shuiwang Ji Texas A&M University 等 10 位作者

原文摘要

The rapid rise in AI conference submissions has driven increasing exploration of large language models (LLMs) for peer review support. However, LLM-based reviewers often generate superficial, formulaic comments lacking substantive, evidence-grounded feedback. We attribute this to the underutilization of two key components of human reviewing: explicit rubrics and contextual grounding in existing work. To address this, we introduce ReviewBench, a benchmark evaluating review text according to paper-specific rubrics derived from official guidelines, the paper’s content, and human-written reviews. We further propose ReviewGrounder, a rubric-guided, tool-integrated multi-agent framework that decomposes reviewing into drafting and grounding stages, enriching shallow drafts via targeted evidence consolidation. Experiments on ReviewBench show that ReviewGrounder, using a Phi-4-14B-based drafter and a GPT-OSS-120B-based grounding stage, consistently outperforms baselines with substantially stronger/larger backbones (e.g., GPT-4.1 and DeepSeek-R1-670B) in both alignment with human judgments and rubric-based review quality across 8 dimensions. The code is available at https://github.com/EigenTom/ReviewGrounder.

cs.CL · cs.MA

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.1894ACL Anthology · ACL 2026

Mind Reader: Latent User Demand-Guided Content Optimization for Generative Search Engine

Tong Chen Beijing University of Aeronautics and Astronautics, JiaWei Guo, Yuxi Li Beihang University, Baiming Chen Beijing University of Aeronautics and Astronautics, Houxing Ren Beijing University of Aeronautics and Astronautics, Zhang Zhiwei, Yunxiang Zhang, Hanyang Xia Beijing University of Aeronautics and Astronautics 等 10 位作者

原文摘要

Generative Search Engines (GSEs) have reshaped information retrieval, and Generative Engine Optimization (GEO) emerges to improve the content visibility in GSEs’ responses. Previous methods mainly rely on empirical strategies or query-dependent preferences of GSEs for content optimization. However, they remain limited in effectiveness as they overlook the latent user search demands in queries that drive content retrieval and response generation of GSEs. To address this, we propose Mind Reader, a novel GEO method to effectively improve the content visibility within the generated responses of GSEs through content optimization guided by the extracted latent demands of user search. Specifically, we propose a decomposition-recombination query augmentation module, which enriches the query with latent semantic information by decomposing it into diverse perspectives, capturing underlying semantic information, and recombining them into variants to support subsequent optimization. Then, we propose a reasoning coverage content optimization module. By optimizing content to cover critical reasoning information of GSEs, we align the content with the user search demands, effectively improving the content visibility. Extensive experiments on widely used GEO-Bench and our proposed PC-GEO show that our method significantly outperforms baselines and effectively improves content visibility (with up to 2.44x objective metrics and 1.23x subjective metrics on average).

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.372ACL Anthology · ACL 2026

On the Emotion Understanding of Synthesized Speech

Yuan Ge, Haishu Zhao Northeastern University, AoKai Hao Northeastern University, Junxiang Zhang, Bei Li Meituan, Xiaoqian Liu Northeastern University, Chenglong Wang Northeastern University, Jianjin Wang 等 13 位作者

原文摘要

Emotion is a core paralinguistic feature in voice interaction. It is widely believed that emotion understanding models learn fundamental representations that transfer to synthesized speech, making emotion understanding results a plausible reward or evaluation metric for assessing emotional expressiveness in speech synthesis. In this work, we critically examine this assumption by systematically evaluating Speech Emotion Recognition (SER) on synthesized speech across datasets, discriminative and generative SER models, and diverse synthesis models. We find that current SER models can not generalize to synthesized speech, largely because speech token prediction during synthesis induces a representation mismatch between synthesized and human speech. Moreover, generative Speech Language Models (SLMs) tend to infer emotion from textual semantics while ignoring paralinguistic cues. Overall, our findings suggest that existing SER models often exploit non-robust shortcuts rather than capturing fundamental features, and paralinguistic understanding in SLMs remains challenging.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-long.505ACL Anthology · ACL 2026

Learning from Evolving Training Dynamics: An Entropy-Maximizing Data Curation Strategy for LLM Supervised Post-Training

Mengxiang Zhang University of Hong Kong, Lingyuan Liu City University of Hong Kong

原文摘要

Supervised post-training is essential for refining Large Language Models (LLMs), yet its effectiveness relies heavily on strategic data curation. Traditional Curriculum Learning (CL) strategies often fail to account for the evolving proficiency of the learner, relying instead on static, single dimensional metrics. We propose EVO-Curate, a dynamic data curation framework that synchronizes sample complexity with the maturing capacity of the LLM. EVO-Curate employs an Adaptive Dynamics Measurer to synthesize instantaneous difficulty and historical variability into a multidimensional utility score. To maintain representational diversity, we introduce an Evolutionary Sampling Scheduler based on an entropy maximizing mechanism. Empirical evaluations across instruction following, mathematical reasoning, and code generation demonstrate that EVO-Curate consistently outperforms standard training baselines and traditional CL methods across various architectures and scales. Specifically, our framework achieves relative performance gains of up to about 10% while maintaining manageable computational overhead. These results establish EVO-Curate as a scalable and model agnostic solution for enhancing the efficiency of modern LLM training pipelines.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表

首次观测 · 记录 #1ACL Anthology · ACL 2026:acl:2026.acl-short.49ACL Anthology · ACL 2026

Skill-Aware Data Selection and Fine-Tuning for Data-Efficient Reasoning Distillation

Lechen Zhang University of Illinois at Urbana-Champaign, Yunxiang Zhang, Wei Hu OpenAI and University of Michigan - Ann Arbor, Lu Wang University of Michigan

原文摘要

Large reasoning models such as DeepSeek-R1 and their distilled variants achieve strong performance on complex reasoning tasks. Yet, distilling these models often demands large-scale data for supervised fine-tuning (SFT), motivating the pursuit of data-efficient training methods. To address this, we propose a skill-centric distillation framework that efficiently transfers reasoning ability to weaker models with two components: (1) Skill-based data selection, which prioritizes examples targeting the student model’s weaker skills, and (2) Skill-aware fine-tuning, which encourages explicit skill decomposition during problem solving. With only 1,000 training examples selected from a 100K teacher-generated corpus, our method surpasses random SFT baselines by +1.6% on Qwen3-4B and +1.4% on Qwen3-8B across five mathematical reasoning benchmarks. Further analysis confirms that these gains concentrate on skills emphasized during training, highlighting the effectiveness of skill-centric training for efficient reasoning distillation.

cs.CL

Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) · 2026 · 正式发表