大模型推理能力 - arXivDaily 专题

2608.07051 2026-08-10 cs.CV 新提交 50%

YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family

YOLO-PEFT：针对YOLO系列的参数高效微调

Xu Lin, WenJie Nie, Jinlong Peng, Weifu Fu, YueXiao Ma, Xiawu Zheng, Yong Liu

专题命中规划推理：planning(abstract)

AI总结 YOLO-PEFT是一种针对YOLO系列的结构感知参数高效微调框架，将适配器部署建模为约束规划问题，在VOC数据集上的实验显示其检测精度优于全微调，且能减少训练内存。

2608.06789 2026-08-10 cs.NI 新提交 50%

EvoRIC: Reinforcement Learning Fine-Tuned LLM-empowered RAN Intelligent Control Toward Autonomous O-RAN

EvoRIC：面向自主O-RAN的、由强化学习微调大语言模型赋能的RAN智能控制器

Lingyan Bao, Jemin Lee, Tony Q. S. Quek

专题命中规划推理：reasoning(abstract)

AI总结针对传统RAN智能算法泛化差、通用LLM算力高且缺领域知识的问题，提出EvoRIC分层框架，结合RLFT与PPO优化LLM，在IAB网络中验证其泛化性与效能，为自主O-RAN提供新方案。

Comments Manuscript submitted 23 April 2026; revised 7 August 2026

2608.06648 2026-08-10 cs.RO 新提交 50%

Plan-and-Avoid: Real-Time Aircraft Trajectory Coordination in a Multi-Agent Environment

规划与避让：多智能体环境下的实时飞行器轨迹协调

Huseyin Emre Tekaslan, Ella M. Atkins, Natasha Neogi

机构 * Virginia Polytechnic Institute and State University（弗吉尼亚理工学院暨州立大学）； NASA Langley Research Center（NASA兰利研究中心）

专题命中规划推理：planning(abstract)

AI总结该研究提出Plan-and-Avoid框架，针对多智能体空域环境，通过实时生成协作建议，在900余起强制着陆案例测试中，以5.7秒最坏响应时间实现优先级轨迹的低延迟安全间隔协调。

2608.06525 2026-08-10 math.GN 新提交 50%

A Precise Treatment of Soft Quotient Topology and Soft Covering Maps

软商拓扑与软覆盖映射的精确处理

Souvik Mandal, Ankur Sarkar

专题命中规划推理：planning(abstract)

AI总结该研究建立软商拓扑基础理论，提出软覆盖映射精确定义，将其应用于多智能体运动规划以降低组合复杂性。

Comments 18 Pages. Comments are welcome

2608.06771 2026-08-10 physics.soc-ph 新提交 50%

Agentic Artificial Intelligence for Reproducible Human-in-the-Loop Environmental Health Research

用于可复现人在环环境健康研究的智能体式人工智能

Edmund Seto

专题命中规划推理：planning(abstract)

AI总结本文提出人在环智能体框架，结合领域知识与编码素养，利用智能体式大语言模型生成R代码，通过美国超级基金场地案例验证其可提升环境健康研究结果的严谨性与可复现性。

Comments 41 pages, 3 figures

2608.06680 2026-08-10 physics.chem-ph 新提交 50%

Physics-Grounded Materials Artificial Intelligence for Reliable Materials Discovery

基于物理学原理的材料人工智能用于可靠的材料发现

Yuhang Wang, Qian Wang, Seong-Hoon Jang, Hao Li

专题命中规划推理：reasoning(abstract)

AI总结该研究提出PhysMat AI框架，整合物理知识用于材料发现，通过多领域示例说明其应用，还规划了从物理感知到自主AI的发展路线，助力可靠材料发现。

2608.04676 2026-08-06 cs.CV 新提交 50%

SurgNarrator: A Generative Retrieval Framework for Surgical Video Understanding

SurgNarrator：面向手术视频理解的生成式检索框架

Yuqing Feng, Jiawei Ma, Kevin Qinghong Lin, Kun Yuan, Nicolas Padoy, Daniel S. Elson, Anh Nguyen, Stamatia Giannarou, Baoru Huang

机构 * University of Liverpool（利物浦大学）； City University of Hong Kong（香港城市大学）； University of Oxford（牛津大学）； University of Strasbourg（斯特拉斯堡大学）； IHU Strasbourg（斯特拉斯堡大学医院研究所）； Imperial College London（帝国理工学院）

专题命中规划推理：reasoning(abstract)

AI总结 SurgNarrator是专为手术视频理解设计的生成式检索框架，通过构建手术词汇表、适配Qwen3-VL-Embedding-8B并采用分层检索策略，在零样本设置下于12个基准上实现性能提升且延迟大幅降低。

Comments This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

2608.04653 2026-08-06 cs.CV cs.RO 新提交 50%

Overcoming Statistical Bias in Action-Controllable World Models

克服动作可控世界模型中的统计偏差

Yuhong Shi, Zhenhao Chu, Jie Wei, Jun Hao, Jianyi Liu, Jingwen Fu

专题命中规划推理：planning(abstract)

AI总结该研究针对动作可控世界模型的统计偏差问题，提出CoCo反事实一致性框架，结合ARC、DE评估指标及Mini-SSMB数据集，在多项任务上提升了动作可控性与视频预测性能。

2608.04634 2026-08-06 cs.MA 新提交 50%

HELENA:Hierarchical Sparse Coordination over a Union of Complementary Topologies for MAS

HELENA：面向多智能体系统的互补拓扑联合上的分层稀疏协调框架

Zhifang Mao, Linyao Zheng, Xuhang Shi, Xiuquan Hou

专题命中规划推理：reasoning(abstract)

AI总结 HELENA 是一种 MAS 框架，通过互补拓扑联合与分层稀疏协调抑制冗余噪声，在八个基准测试上实现最优结果，平均提升 3.47%，MMLU-Pro 最高提升 10.34%，难基准提升更显著且附加成本合理。

2608.04416 2026-08-06 cs.HC 新提交 50%

Preference-Driven Online Adaptation for Personalized Interaction Initiation in Proactive AI Assistants

主动式AI助手中个性化交互发起的偏好驱动在线适应

Yufeng Wang, Wei Zhang, Zhiquan Wen, Jinwu Hu, Linhui Xiao, Tianlu Pan, Qingfang Zheng, Mingkui Tan

专题命中规划推理：reasoning(abstract)

AI总结本文针对主动式AI助手个性化交互时机难确定的问题，提出EOPA方法，在ProPerSim基准上较最强基线提升F1分数19.80个百分点，降低推理延迟与平均每日适应时间。

2608.03648 2026-08-05 cs.MA 新提交 50%

Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate

群体视角至关重要：调控辩论关系可缓解多智能体辩论中的盲目从众

Hao Wu, Shoucheng Song, Chang Yao, Haoyu Wang, Huaiyu Wan, Youfang Lin, Kai Lv

专题命中规划推理：reasoning(abstract)

AI总结针对多智能体辩论中LLMs易盲目从众的问题，提出DEAR框架，通过动态调控辩论关系缓解该问题，实验显示其性能更优且token消耗显著降低。

2608.03631 2026-08-05 cs.CV 新提交 50%

SEER: A Self-Grounded Evidence Interface for Controlled Spatial Relation Classification

SEER：用于受控空间关系分类的自 grounding 证据接口

Feixiang Liu, Likun Wang, Qiang Qiu, Hui Xu, Huawei Shen, Xueqi Cheng

专题命中规划推理：reasoning(abstract)

AI总结本文提出针对冻结VLM的无训练推理时证据接口SEER，通过构建查询特定证据缓解空间关系分类错误，在多数据集上取得显著性能提升，证明该干预措施的有效性。

Comments 23 pages total, 2 figures. Code: https://github.com/SouthWinter/SEER

2608.01949 2026-08-04 cs.IR 新提交 50%

A Self-Triggered Agentic Push Recommendation System

一种自触发智能推送推荐系统

Zhao-Yu Zhang, Qingying Chen, Chunyuan Zheng, Jing Zhou, Jian Sun, Siqi Chen, Leiying Chen, Chuan Zhou, Huiyou Jiang, Xin Tao, Haoxuan Li, Zhouchen Lin

专题命中规划推理：planning(abstract)

AI总结本文提出端到端自触发智能推送系统STEPS，含规划、执行、过滤智能体，已在抖音部署，可提升用户活跃天数、降低权限禁用率并减少计算开销。

2608.01561 2026-08-04 cs.HC 新提交 50%

FedWorld: Scope-Aware Federation of Agent World Models

FedWorld：感知范围的智能体世界模型联邦

Yuchao Hou

专题命中规划推理：planning(abstract)

AI总结 FEDWORLD是一种感知范围的智能体世界模型联邦协议，通过交换结构化抽象转移规则，减少冲突动态下的负迁移，提升智能体任务成功率。

2608.01250 2026-08-04 eess.SY cs.SY 新提交 50%

Smoothing the Ramp, Not the Peak: Scheduling-Induced Power Dynamics of LLM Inference and Their Grid-Scale Consequences

平滑斜坡而非峰值：LLM推理的调度诱导功率动态及其电网级后果

Pan Li, Yize Chen, Xia Miao, Dai Wang

专题命中规划推理：planning(abstract)

AI总结该研究揭示LLM分块预填充调度可降低功率斜坡速率，随系统饱和度提升效益增大，经转化为电网调节备用采购问题后，可减少电网快速斜坡备用容量，为运营商提供无成本需求侧塑形工具。

Comments 15 pages, 9 figures

2608.00959 2026-08-04 cs.GT cs.RO cs.SY eess.SY 新提交 50%

MixedComplementarityProblems.jl: A Fast, Batched, Open-Source Interior Point Solver for Mixed Complementarity Problems

MixedComplementarityProblems.jl：用于混合互补问题的快速、批处理开源内点求解器

David Fridovich-Keil

专题命中规划推理：planning(abstract)

AI总结本文提出开源Julia实现的混合互补问题内点求解器MixedComplementarityProblems.jl，其批处理CPU版本处理多智能体变道轨迹博弈的速度约为PATH的100倍，GPU版本仅在KKT系统规模大时更具优势，且可靠性与PATH相当。

2608.00690 2026-08-04 cs.NI 新提交 50%

LLM-Assisted Coalition Formation for Cooperative Perception in Autonomous Driving

大语言模型辅助的自动驾驶协同感知联盟形成

Ahmad Sarlak, Hao Wang, Rahul Amin, Abolfazl Razi

专题命中规划推理：reasoning(abstract)

AI总结该研究针对现有协同感知方法未优化车辆选择的问题，提出LLM辅助的联盟形成框架，结合DPP与ADMM优化，在OPV2V等数据集上实现更优性能与网络效率。

Comments Accepted for presentation at IEEE Global Communications Conference (GLOBECOM 2026), Cognitive Radio and AI-Enabled Networks Symposium

2608.00669 2026-08-04 cs.IR 新提交 50%

GARDRec: Decision-Level Graph Grounding for Large Language Model Recommendation

GARDRec：面向大语言模型推荐的决策级图接地方法

Yong Wang, Hongliang Sun, Jinlan Liu, Hua Zhang, Dianbo Sui, Dianhui Chu, Zhiying Tu

专题命中规划推理：reasoning(abstract)

AI总结本文针对现有图增强型LLM推荐器排序决策受用户-物品关系约束弱的问题，提出GARDRec框架，通过图接地技术提升下一项推荐的候选排序性能，经实验验证其优于代表性基线。

Comments 18 pages, 2 figures

2608.00648 2026-08-04 cs.MA 新提交 50%

MDGAM-Based Cooperative Task Scheduling for Communication-Constrained Distributed Multi-Agent Systems

基于MDGAM的通信受限分布式多智能体系统协同任务调度

Licheng Wang, Mingtao Huang, Yuan Shen

专题命中规划推理：planning(abstract)

AI总结针对通信受限分布式多智能体系统的任务调度问题，本文提出基于MDGAM的神经调度框架，结合GRMAPG算法，经实验验证其可提升任务完成性能。

2607.29310 2026-08-03 cs.CV 新提交 50%

CALM-AH: An ABAW11-Calibrated Multimodal Ensemble with Reliability-Gated Multi-Expert Consensus for Video-Level Ambivalence and Hesitancy Recognition

CALM-AH：面向视频层面矛盾与犹豫识别的ABAW11校准多模态集成模型，采用可靠性门控多专家共识机制

Wenzhuo Sun, Mingjian Liang, Richard Attfield, Zongyuan Ge, Xuelian Cheng, Pamela Carreno-Medrano

机构 * Monash University（莫纳什大学）； Monash Suzhou Research Institute（莫纳什苏州研究院）

专题命中规划推理：verifier(abstract)

AI总结针对ABAW11视频层面矛盾与犹豫识别任务，提出CALM-AH多模态集成模型及RG-MEC门控多专家共识机制，在ABAW11数据集上分别取得0.7525与0.7771的Macro-F1性能。

2607.29175 2026-08-03 cs.SE 新提交 50%

Execution-First Synthetic Tool-Use Trace Generation for LLM Agents

面向LLM智能体的优先执行式合成工具使用轨迹生成

Hafsa Ouajdi, Francesco Giannuzzo, Alaa Boukhary, Paolo Papotti, Gerard Conangla, Adam Elwood

专题命中规划推理：reasoning(abstract)

AI总结针对传统数据合成的局限，提出优先执行式框架SyntheticAgentTraceQA生成工具增强型智能体的监督数据，经四工具生态系统及Qwen模型验证，该框架可提升工具执行等性能，且揭示了掩码与全监督的权衡。

2607.28822 2026-08-03 eess.SY cs.SY 新提交 50%

Agentic Artificial Intelligence for Power Systems: Strategies to Identify and Close Capability Gaps

面向电力系统的智能体人工智能：识别与缩小能力差距的策略

Eve Tsybina, Samim Konjicija, Slaven Peles

专题命中规划推理：planning(abstract)

AI总结该研究针对电力系统规划任务，复现当前最优Agentic AI并评估其能力，发现其仅能解决部分低复杂度问题，明确了能力升级需求，强调需采用严格测试与可复现基准评估Agentic AI进展。

Comments 5 pages, 1 figure, 2 tables

2607.28601 2026-07-31 cs.MA q-bio.NC 新提交 50%

Using Theory of Mind to Arbitrate between Social and Non-social Learning

运用心智理论在社会学习与非社会学习之间进行仲裁

Lance Ying, Ryan Truong, Joshua B. Tenenbaum, Samuel J. Gershman

专题命中规划推理：reasoning(abstract)

AI总结该研究提出理性心智化模型，通过推理智能体目标与行动信息性权衡社会、非社会学习的效用，用新游戏验证其可定量捕捉人类学习策略的权衡，揭示选择性社会学习受心智理论指导以实现效用最大化。

Comments 35 pages, includes supplementary information

2607.28341 2026-07-31 cs.CV 新提交 50%

Capturing Token Tendencies for Training-Free Token Pruning in Multimodal Large Language Models

捕捉令牌趋势以实现多模态大语言模型中无需训练的令牌剪枝

Jie Ma, Zhike Qiu, Jie Gao, Jiayi Ji, Qian Chen, Xiaoshuai Sun, Rongrong Ji

机构 * Xiamen University（厦门大学）； Xiamen Ocean Vocational College（厦门海洋职业技术学院）； Sino-Russian Research Center for Digital Economy（中俄数字经济研究中心）

专题命中规划推理：reasoning(abstract)

AI总结该研究针对多模态大语言模型现有无训练剪枝方法忽略令牌重要性动态变化的问题，提出趋势感知剪枝框架，通过捕捉注意力流动量实现动态修正，大幅减少视觉令牌且保持性能，实现高效多模态推理。

2607.28186 2026-07-31 cs.CV 新提交 50%

Think with Extra-Image: A Farmland Segmentation Agent Driven by Spatio-Temporal Information Gain

借助图像外信息思考：由时空信息增益驱动的农田分割智能体

Haiyang Wu, Weiliang Mu, Zhuofei Du, Dandan Zhong, Kaijie Shi, Haifeng Li, Chao Tao

专题命中规划推理：reasoning(abstract)

AI总结该研究针对现有农田遥感图像分割范式的不足，提出动态分割智能体FarmSeeker，构建支持推理查询的全球高分辨率基准GSFS-Bench，其分割性能比现有方法更稳定。

2607.27435 2026-07-31 cs.DB 新提交 50%

AgenticER: the next frontier in Entity Resolution

AgenticER：实体解析的下一个前沿领域

George Papadakis, Panos Korovesis, Manolis Koubarakis, Themis Palpanas

专题命中规划推理：reasoning(abstract)

AI总结针对现有实体解析被动范式无法适配异构流数据场景的问题，提出Agentic ER新范式，将其建模为自主智能体的序列决策过程，构建参考架构并明确研究挑战，开拓数据管理与智能体交叉的新研究方向。

Comments Vision paper submitted to VLDB, 6 pages

2607.27100 2026-07-30 cs.CY 新提交 50%

Can Large Language Models Represent Urban Publics? Behavioral Replication and Population Mismatch in an Affordable-Housing Experiment

大语言模型能否代表城市公众？经济适用房实验中的行为复制与人口匹配问题

Yuxuan Cai, Yequan Hu, Hongqian Li, Zhanghong Ju, Shuying Guo

专题命中规划推理：planning(abstract)

AI总结该研究以美国经济适用房调查实验对比8款开放权重LLMs与843名受访者，发现LLMs虽能近似部分总体对比，却无法保留人口结构、组内异质性等关键特征，城市规划的模型评估需测试空间与社会结构的模拟保留情况。

Comments Yuxuan Cai and Yequan Hu contributed equally to this work

2607.26893 2026-07-30 cs.IR 新提交 50%

Beyond Action Imitation: Learning a Decision-Aware User Simulator for Online Advertising

超越动作模仿：面向在线广告的决策感知用户模拟器学习

Zipeng Chen, Jiaer Zheng, Xiangyang Xu, Xinyu Lin, Zhaobin Wang, Zhaohui Liu, Qianjin Xiang, Xiaoyu Zhao, Zhuozhen Yu, Guangshuo Wang, Daxing Chen, Junwei Pan, Zhangbin Zhu, Chengguo Yin, Hao Chen, Tat-Seng Chua, Haijie Gu, Jie Jiang

专题命中规划推理：reasoning(abstract)

AI总结针对现有用户模拟器仅基于单领域交互模仿动作的缺陷，本文提出决策感知用户模拟器DASH，通过上下文工程、蒸馏思考轨迹与定制规则奖励模型，在腾讯广告数据上验证了其多维度优势。

Comments 23 pages, 10 figures, 9 tables

2607.26836 2026-07-30 cs.CR 新提交 50%

Before Agents Speak: Pre-hoc Failure Risk Inference in Multi-Agent Systems

智能体交流前：多智能体系统中的事前失败风险推断

Shi Lin, Chenpei Wang, Peng Qian, Dezhang Kong, Minghao Li, Yufeng Li, Xun Wang

专题命中规划推理：reasoning(abstract)

AI总结本研究针对多智能体系统中幻觉级联失败的问题，提出事前风险推断框架HalluProp，通过建模内在幻觉风险与传播机制实现早期诊断，性能优于事后方法，可提升系统可靠性。

2607.25124 2026-07-29 cs.HC 新提交 50%

UrbanTrace: LLM-Assisted Discovery and Semantics-Aware Integration of Spatial Data

UrbanTrace：基于大语言模型的空间数据发现与语义感知集成

Sonia Castelo, Eden Wu, Joao Rulff, Harish Doraiswamy, Juliana Freire, Claudio Silva

专题命中规划推理：reasoning(abstract)

AI总结研究针对城市决策中异构空间数据整合问题，提出UrbanTrace视觉分析系统，利用离线分析器结合大语言模型，通过三个交互式视图实现有效空间聚合，经评估在数据发现上表现优异，将空间聚合敏感性转化为视觉资产。

Comments 11 pages, 7 figures. Accepted to IEEE VIS 2026 (Full Papers Track). Author's version accepted for publication in IEEE Transactions on Visualization and Computer Graphics

1. 规划推理 920 篇

YOLO-PEFT: Parameter-Efficient Fine-Tuning on YOLO Family

EvoRIC: Reinforcement Learning Fine-Tuned LLM-empowered RAN Intelligent Control Toward Autonomous O-RAN

Plan-and-Avoid: Real-Time Aircraft Trajectory Coordination in a Multi-Agent Environment

A Precise Treatment of Soft Quotient Topology and Soft Covering Maps

Agentic Artificial Intelligence for Reproducible Human-in-the-Loop Environmental Health Research

Physics-Grounded Materials Artificial Intelligence for Reliable Materials Discovery

SurgNarrator: A Generative Retrieval Framework for Surgical Video Understanding

Overcoming Statistical Bias in Action-Controllable World Models

HELENA:Hierarchical Sparse Coordination over a Union of Complementary Topologies for MAS

Preference-Driven Online Adaptation for Personalized Interaction Initiation in Proactive AI Assistants

Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate

SEER: A Self-Grounded Evidence Interface for Controlled Spatial Relation Classification

A Self-Triggered Agentic Push Recommendation System

FedWorld: Scope-Aware Federation of Agent World Models

Smoothing the Ramp, Not the Peak: Scheduling-Induced Power Dynamics of LLM Inference and Their Grid-Scale Consequences

MixedComplementarityProblems.jl: A Fast, Batched, Open-Source Interior Point Solver for Mixed Complementarity Problems

LLM-Assisted Coalition Formation for Cooperative Perception in Autonomous Driving

GARDRec: Decision-Level Graph Grounding for Large Language Model Recommendation

MDGAM-Based Cooperative Task Scheduling for Communication-Constrained Distributed Multi-Agent Systems

CALM-AH: An ABAW11-Calibrated Multimodal Ensemble with Reliability-Gated Multi-Expert Consensus for Video-Level Ambivalence and Hesitancy Recognition

Execution-First Synthetic Tool-Use Trace Generation for LLM Agents

Agentic Artificial Intelligence for Power Systems: Strategies to Identify and Close Capability Gaps

Using Theory of Mind to Arbitrate between Social and Non-social Learning

Capturing Token Tendencies for Training-Free Token Pruning in Multimodal Large Language Models

Think with Extra-Image: A Farmland Segmentation Agent Driven by Spatio-Temporal Information Gain

AgenticER: the next frontier in Entity Resolution

Can Large Language Models Represent Urban Publics? Behavioral Replication and Population Mismatch in an Affordable-Housing Experiment

Beyond Action Imitation: Learning a Decision-Aware User Simulator for Online Advertising

Before Agents Speak: Pre-hoc Failure Risk Inference in Multi-Agent Systems

UrbanTrace: LLM-Assisted Discovery and Semantics-Aware Integration of Spatial Data