大模型对齐与安全 - arXivDaily 专题

2502.10844 2026-02-03 cs.HC 50%

Be Friendly, Not Friends: How LLM Sycophancy Shapes User Trust

友好而非朋友：LLM谄媚如何塑造用户信任

Yuan Sun, Ting Wang

专题命中幻觉与事实性：trustworthy(abstract)

AI总结研究探讨LLM谄媚如何影响用户信任，通过实验发现适应性对话态度影响用户对LLM真实性和信任度的感知。

2602.01264 2026-02-03 cs.HC 50%

Shades of Uncertainty: How AI Uncertainty Visualizations Affect Trust in Alzheimer's Predictions

不确定性之影：AI不确定性可视化如何影响阿尔茨海默病预测的可信度

Jonatan Reyes, Mina Massoumi, Anil Ufuk Batmaz, Marta Kersten-Oertel

专题命中幻觉与事实性：trustworthy(abstract)

AI总结本研究探讨了AI不确定性可视化对阿尔茨海默病预测可信度的影响，发现连续编码提升可靠性，而二元编码增强即时信心，揭示了高不确定性下专业背景的权衡。

2601.23041 2026-02-02 cs.CV 50%

One-shot Optimized Steering Vector for Hallucination Mitigation for VLMs

单次优化的转向向量用于VLMs的幻觉缓解

Youxu Shi, Suorong Yang, Dong Liu

机构 * University of Science（科学大学）； Nanjing University（南京大学）

专题命中幻觉与事实性：safety(abstract)

AI总结本文提出OSGA框架，通过单次优化生成通用转向向量，有效缓解VLMs的幻觉问题并提升安全性。

2601.22099 2026-01-30 physics.soc-ph 50%

Towards Universal Urban Patterns-of-Life Simulation

迈向通用的城市生活方式模拟

Sandro M. Reia, Henrique F. de Arruda, Shiyang Ruan, Taylor Anderson, Hamdi Kavak, Dieter Pfoser

专题命中幻觉与事实性：alignment(abstract)

AI总结本文提出了一种可扩展且通用的城市生活方式模拟框架，通过模拟日常活动生成模式，并验证了其在不同城市中的有效性，适用于城市系统中的多种场景分析。

2601.21126 2026-01-30 cs.MA cs.RO 50%

AI-Augmented Density-Driven Optimal Control (D2OC) for Decentralized Environmental Mapping

AI增强的密度驱动最优控制（D2OC）用于去中心化环境映射

Kooktae Lee, Julian Martinez

机构 * Department of Mechanical Engineering, New Mexico Institute of Mining and Technology（机械工程系，新墨西哥矿业与技术研究院）

专题命中幻觉与事实性：alignment(abstract)

AI总结本文提出AI增强的密度驱动最优控制方法，用于在传感和通信受限条件下实现多智能体环境映射，通过自适应机制提升密度估计精度和鲁棒性。

2510.26441 2026-01-27 cs.CV 50%

A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models

A-TPT：面向视觉语言模型测试时提示微调的角多样性校准特性

Shihab Aaqil Ahamed, Udaya S. K. P. Miriya Thanthrige, Ranga Rodrigo, Muhammad Haris Khan

机构 * Dept. of Electronic and Telecommunication Engineering, University of Moratuwa（摩图瓦大学电子与电信工程系）； Mohamed bin Zayed University of Artificial Intelligence（穆罕默德·本·扎耶德人工智能大学）

专题命中幻觉与事实性：safety(abstract)

AI总结 A-TPT通过引入角度多样性提升视觉语言模型测试时提示微调的校准性能，有效减少校准误差并提升适应能力。

Comments Accepted at ICLR 2026

2410.17406 2026-01-26 cs.CR cs.IR 50%

ProveRAG: Provenance-Driven Vulnerability Analysis with Automated Retrieval-Augmented LLMs

ProveRAG: 基于溯源的漏洞分析与自动化检索增强大语言模型

Reza Fayyazi, Stella Hoyos Trueba, Michael Zuzak, Shanchieh Jay Yang

专题命中幻觉与事实性：alignment(abstract)

AI总结 ProveRAG通过自动化检索增强和自我批评机制，提升大语言模型在漏洞分析中的准确性和时效性，提供高精度的漏洞利用和缓解策略。

Journal ref IEEE Access, vol. 13, pp. 212815-212826, 2025

2601.15707 2026-01-23 cs.RO 50%

D-Optimality-Guided Reinforcement Learning for Efficient Open-Loop Calibration of a 3-DOF Ankle Rehabilitation Robot

基于D-最优性的强化学习用于3-自由度踝关节康复机器人高效开环校准

Qifan Hu, Branko Celler, Weidong Mu, Steven W. Su

机构 * Affiliated Provincial Hospital, Shandong First Medical University（山东第一医科大学附属省立医院）； Faculty of Engineering, University of New South Wales（新南威尔士大学工程学院）； Faculty of Engineering and IT, University of Technology Sydney（悉尼大学技术与信息工程学院）

专题命中幻觉与事实性：alignment(abstract)

AI总结本文提出基于D-最优性的强化学习方法，用于高效校准三自由度踝关节康复机器人，提升校准效率和参数估计的鲁棒性。

2509.25856 2026-01-23 cs.CV 50%

PatchEAD: Unifying Industrial Visual Prompting Frameworks for Patch-Exclusive Anomaly Detection

PatchEAD: 统一工业视觉提示框架用于补丁专属异常检测

Po-Han Huang, Jeng-Lin Li, Po-Hsuan Huang, Ming-Ching Chang, Wei-Chao Chen

机构 * Inventec Corporation（Inventec公司）； University at Albany, State University of New York（纽约州立大学阿尔巴尼分校）

专题命中幻觉与事实性：alignment(abstract)

AI总结 PatchEAD提出统一的补丁聚焦框架，实现无需训练的工业异常检测，兼容多种基础模型并提升补丁相似性鲁棒性。

Comments 10 pages, 5 figures. WACV 2026 (Accepted)

2601.07761 2026-01-13 cs.CV 50%

Video Evidence to Reasoning Efficient Video Understanding via Explicit Evidence Grounding

视频证据推理：通过显式证据接地实现高效的视频理解

Yanxiang Huang, Guohua Gao, Zhaoyang Wei, Jianyuan Ni

机构 * Department of Applied Mathematics, The Hong Kong Polytechnic University（应用数学系，香港理工大学）

专题命中幻觉与事实性：alignment(abstract)

AI总结本文提出CoE框架，通过显式证据接地模块和强化学习优化的锚定协议，提升视频推理的准确性和可靠性。

Comments 6 pages

Journal ref ICME 2026

2601.06798 2026-01-13 cs.IR 50%

Unleashing the Native Recommendation Potential: LLM-Based Generative Recommendation via Structured Term Identifiers

释放原生推荐潜力：基于结构化术语标识符的LLM生成推荐

Zhiyang Zhang, Junda She, Kuo Cai, Bo Chen, Shiyao Wang, Xinchen Luo, Qiang Luo, Ruiming Tang, Han Li, Kun Gai, Guorui Zhou

专题命中幻觉与事实性：alignment(abstract)

AI总结本文提出GRLM框架，通过结构化术语标识符提升生成推荐系统的性能和通用性。

2601.05721 2026-01-12 cs.SE 50%

From Issues to Insights: RAG-based Explanation Generation from Software Engineering Artifacts

从问题到洞察：基于RAG的软件工程制品解释生成

Daniel Pöttgen, Mersedeh Sadeghi, Max Unterbusch, Andreas Vogelsang

专题命中幻觉与事实性：alignment(abstract)

AI总结本文首次应用RAG方法从问题跟踪数据生成解释，实现90%的人工解释一致性，展示了结构化问题数据在提升软件系统可解释性方面的潜力。

Comments Accepted at NLBSE 2026, Rio de Janeiro, Brazil

2509.24307 2026-01-09 cs.HC 50%

Exploring Similarity between Neural and LLM Trajectories in Language Processing

探索神经与大语言模型在语言处理中的相似性

Xin Xiao, Kaiwen Wei, Jiang Zhong, Xuekai Wei, Mingliang Zhou

专题命中幻觉与事实性：alignment(abstract)

AI总结本研究通过比较LLMs与人类大脑在语言处理中的轨迹相似性，揭示了语义整合机制和实时处理动态的差异。

2601.02766 2026-01-07 cs.RO cs.AR 50%

Advancing Assistive Robotics: Multi-Modal Navigation and Biophysical Monitoring for Next-Generation Wheelchairs

推动辅助机器人：多模态导航与生物物理监测用于下一代轮椅

Md. Anowar Hossain, Mohd. Ehsanul Hoque

专题命中幻觉与事实性：safety(abstract)

AI总结本文提出了一种多模态轮椅控制系统，结合多种输入接口与生物物理监测，提升患者独立性并实现护理人员实时监督。

2503.06472 2026-01-07 cs.CV cs.MM 50%

CalliReader: Contextualizing Chinese Calligraphy via an Embedding-Aligned Vision-Language Model

CalliReader：通过嵌入对齐的视觉-语言模型 contextualizing 中文书法

Yuxuan Luo, Jiaqi Tang, Chenyi Huang, Feiyang Hao, Zhouhui Lian

机构 * Wangxuan Institute of Computer Technology, Peking University（北京大学计算机技术研究院）

专题命中幻觉与事实性：alignment(abstract)

AI总结 CalliReader通过字符级切片、视觉-文本对齐和嵌入指令微调，解决中文书法的上下文识别问题，实现比现有方法和专业书法家更高的准确性与更低的幻觉率。

Comments 11 pages

Journal ref Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) 2025

2512.18671 2025-12-23 cs.CV 50%

SmartSight: Mitigating Hallucination in Video-LLMs Without Compromising Video Understanding via Temporal Attention Collapse

SmartSight: 通过时间注意力崩溃缓解视频大语言模型中的幻觉而不影响视频理解

Yiming Sun, Mi Zhang, Feifei Li, Geng Hong, Min Yang

专题命中幻觉与事实性：safety(abstract)

AI总结 SmartSight通过时间注意力崩溃技术，在不牺牲视频理解能力的前提下，有效降低视频大语言模型的幻觉问题。

Comments AAAI26 accepted

2508.17201 2025-12-19 physics.med-ph 50%

Root Cause Analysis of Radiation Oncology Incidents Using Large Language Models

利用大语言模型进行放射肿瘤事件的根本原因分析

Yuntao Wang, Mariluz De Ornelas, Matthew T. Studenski, Elizabeth Bossart, Siamak P. Najad-Davarani, Yunze Yang

专题命中幻觉与事实性：safety(abstract)

AI总结利用大语言模型分析放射肿瘤事件的根本原因，评估其在提升患者安全中的应用潜力。

2512.11096 2025-12-15 cs.HC 50%

Your plan may succeed, but what about failure? Investigating how people use ChatGPT for long-term life task planning

你的计划可能成功，但失败呢？探究人们如何利用ChatGPT进行长期生活任务规划

Ben Wang, Jiqun Liu

专题命中幻觉与事实性：trustworthy(abstract)

AI总结研究探讨人们如何利用ChatGPT进行长期生活任务规划，发现AI在结构化目标和生成想法方面有帮助，但输出缺乏个性化和适应性，需用户主动调整。

2512.09249 2025-12-11 cond-mat.mtrl-sci 50%

Auto-3DPFM: Automating Polarization-Vector Mapping at the Nanoscale

Auto-3DPFM：在纳米尺度上自动化极化矢量映射

Ralph Bulanadi, Marti Checa, Michelle Wang, Franck Rothen, John Lasseter, Sumner B. Harris, Daniel Sando, Valanoor Nagarajan, Liam Collins, Stephen Jesse, Rama Vasudevan, Yongtao Liu

专题命中幻觉与事实性：alignment(abstract)

AI总结 Auto-3DPFM通过自动化技术实现纳米尺度极化矢量的高精度表征，提升铁电材料研究的效率与准确性。

2508.04679 2025-12-11 cs.HC 50%

MisVisFix: An Interactive Dashboard for Detecting, Explaining, and Correcting Misleading Visualizations using Large Language Models

MisVisFix: 一种利用大语言模型检测、解释和纠正误导性可视化信息的交互式仪表板

Amit Kumar Das, Klaus Mueller

专题命中幻觉与事实性：trustworthy(abstract)

AI总结 MisVisFix利用大语言模型提供交互式工具，用于检测、解释和纠正误导性可视化信息，提升数据解读的准确性和可信度。

Comments 11 pages, 6 figures. Accepted at IEEE VIS: Visualization & Visual Analytics 2025 conference, November 2-7, 2025, Vienna, Austria

Journal ref IEEE Transactions on Visualization and Computer Graphics (TVCG), PrePrints 5555, pp. 1-11, 2025

2511.22019 2025-12-09 cs.CV 50%

Intra-Class Probabilistic Embeddings for Uncertainty Estimation in Vision-Language Models

类内概率嵌入用于视觉-语言模型中的不确定性估计

Zhenxiang Lin, Maryam Haghighat, Will Browne, Dimity Miller

机构 * Queensland University of Technology（昆士兰理工大学）

专题命中幻觉与事实性：safety(abstract)

AI总结本研究提出一种无需训练的后处理方法，通过类内概率嵌入提升视觉-语言模型的不确定性估计，有效检测错误预测。

Comments Accepted at the IEEE/CVF Winter Conference on Applications of Computer Vision 2026

2510.22171 2025-12-02 cs.CV 50%

HARMONY: Hidden Activation Representations and Model Output-Aware Uncertainty Estimation for Vision-Language Models

HARMONY：隐藏的激活表示和模型输出感知的不确定性估计用于视觉-语言模型

Erum Mushtaq, Zalan Fabian, Yavuz Faruk Bakman, Anil Ramakrishna, Mahdi Soltanolkotabi, Salman Avestimehr

机构 * University of Southern California（南加州大学）； Amazon AGI（亚马逊人工智能实验室）

专题命中幻觉与事实性：alignment(abstract)

AI总结 HARMONY通过整合生成token、模型输出不确定性分数和隐藏表示，提升视觉-语言模型的不确定性估计性能。

2511.22746 2025-12-01 cs.HC 50%

Epistemic Fragility in Large Language Models: Prompt Framing Systematically Modulates Misinformation Correction

大型语言模型中的知识脆弱性：提示框架系统性地调节 misinformation 的纠正

Sekoul Krastev, Hilary Sweatman, Anni Sternisko, Steve Rathje

专题命中幻觉与事实性：alignment(abstract)

AI总结研究发现提示框架对大型语言模型纠正虚假信息的能力有显著影响，揭示了模型在知识准确性上的脆弱性。

2511.19806 2025-11-26 cs.CV 50%

Reading Between the Lines: Abstaining from VLM-Generated OCR Errors via Latent Representation Probes

在行之间阅读：通过潜在表示探测避免VLM生成的OCR错误

Jihan Yao, Achin Kulshrestha, Nathalie Rauschmayr, Reed Roberts, Banghua Zhu, Yulia Tsvetkov, Federico Tombari

机构 * University of Washington（华盛顿大学）； Google（谷歌）

专题命中幻觉与事实性：safety(abstract)

AI总结本文提出通过潜在表示探测提升VLM在OCR任务中的回避准确性，通过内部状态检测置信度信号，提高系统可靠性。

2511.13533 2025-11-25 cs.CV 50%

Minimax Multi-Target Conformal Prediction with Applications to Imaging Inverse Problems

最小最大多目标符合预测及其在成像反问题中的应用

Jeffrey Wen, Rizwan Ahmad, Philip Schniter

机构 * Department of Electrical and Computer Engineering（电气与计算机工程系）； The Ohio State University（俄亥俄州立大学）； Department of Biomedical Engineering（生物医学工程系）

专题命中幻觉与事实性：safety(abstract)

AI总结本文提出了一种渐近最小最大方法用于多目标符合预测，应用于多指标盲图像质量评估、多任务不确定性量化和多轮测量获取，通过合成和MRI数据验证了其有效性。

Journal ref Transactions on Machine Learning Research, 11/2025. https://openreview.net/forum?id=53FEYwDQK0

2511.17875 2025-11-25 math.OC 50%

Modeling and Calibration of Supplier Selection Problem in Freight Agent-Based Simulations

货运 agent 仿真实验中供应商选择问题的建模与校准

Abdelrahman Ismael, Taner Cokyasar

专题命中幻觉与事实性：alignment(abstract)

AI总结本研究提出了一种大规模校准的 agent 仿真实验模型，用于建模和校准供应商选择问题，以提高货运网络流的预测精度和供应链的韧性。

2511.11984 2025-11-18 cs.CV 50%

From Classification to Cross-Modal Understanding: Leveraging Vision-Language Models for Fine-Grained Renal Pathology

Zhenhao Guo, Rachit Saluja, Tianyuan Yao, Quan Liu, Junchao Zhu, Haibo Wang, Daniel Reisenbüchler, Yuankai Huo, Benjamin Liechty, David J. Pisapia, Kenji Ikemura, Steven Salvatoree, Surya Seshane, Mert R. Sabuncu, Yihe Yang, Ruining Deng

机构 * New York University（纽约大学）； Cornell Tech（康奈尔科技）； Vanderbilt University（范德比大学）； Carnegie Mellon University（卡内基梅隆大学）； University of Regensburg（莱茵河畔大学）； Weill Cornell Medicine（韦尔·科恩医学中心）； Northwell Health（北well健康）

专题命中幻觉与事实性：alignment(abstract)

2511.11106 2025-11-17 cs.MM cs.CV cs.SD 50%

AccKV: Towards Efficient Audio-Video LLMs Inference via Adaptive-Focusing and Cross-Calibration KV Cache Optimization

Zhonghua Jiang, Kui Chen, Kunxi Li, Keting Yin, Yiyun Zhou, Zhaode Wang, Chengfei Lv, Shengyu Zhang

专题命中幻觉与事实性：alignment(abstract)

2511.10974 2025-11-17 cs.CV 50%

Preserving Cross-Modal Consistency for CLIP-based Class-Incremental Learning

Haoran Chen, Houze Xu, Micah Goldblum, Daoguo Dong, Zuxuan Wu

机构 * Institute of Trustworthy Embodied AI（可信具身人工智能研究院）； Fudan University（复旦大学）； Shanghai Collaborative Innovation Center of Intelligent Visual Computing（上海智能视觉计算协同创新中心）； Columbia University（哥伦比亚大学）

专题命中幻觉与事实性：alignment(abstract)

2507.19110 2025-11-14 cs.CV 50%

LISA: A Layer-wise Integration and Suppression Approach for Hallucination Mitigation in Multimodal Large Language Models

Zhihui Guo, Xin Man, Hui Xu, Jie Shao, Zhiguo Jiang, Xianchao Zhang, Heng Tao Shen

专题命中幻觉与事实性：alignment(abstract)

1. 幻觉与事实性 1740 篇

Be Friendly, Not Friends: How LLM Sycophancy Shapes User Trust

Shades of Uncertainty: How AI Uncertainty Visualizations Affect Trust in Alzheimer's Predictions

One-shot Optimized Steering Vector for Hallucination Mitigation for VLMs

Towards Universal Urban Patterns-of-Life Simulation

AI-Augmented Density-Driven Optimal Control (D2OC) for Decentralized Environmental Mapping

A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-Language Models

ProveRAG: Provenance-Driven Vulnerability Analysis with Automated Retrieval-Augmented LLMs

D-Optimality-Guided Reinforcement Learning for Efficient Open-Loop Calibration of a 3-DOF Ankle Rehabilitation Robot

PatchEAD: Unifying Industrial Visual Prompting Frameworks for Patch-Exclusive Anomaly Detection

Video Evidence to Reasoning Efficient Video Understanding via Explicit Evidence Grounding

Unleashing the Native Recommendation Potential: LLM-Based Generative Recommendation via Structured Term Identifiers

From Issues to Insights: RAG-based Explanation Generation from Software Engineering Artifacts

Exploring Similarity between Neural and LLM Trajectories in Language Processing

Advancing Assistive Robotics: Multi-Modal Navigation and Biophysical Monitoring for Next-Generation Wheelchairs

CalliReader: Contextualizing Chinese Calligraphy via an Embedding-Aligned Vision-Language Model

SmartSight: Mitigating Hallucination in Video-LLMs Without Compromising Video Understanding via Temporal Attention Collapse

Root Cause Analysis of Radiation Oncology Incidents Using Large Language Models

Your plan may succeed, but what about failure? Investigating how people use ChatGPT for long-term life task planning

Auto-3DPFM: Automating Polarization-Vector Mapping at the Nanoscale

MisVisFix: An Interactive Dashboard for Detecting, Explaining, and Correcting Misleading Visualizations using Large Language Models

Intra-Class Probabilistic Embeddings for Uncertainty Estimation in Vision-Language Models

HARMONY: Hidden Activation Representations and Model Output-Aware Uncertainty Estimation for Vision-Language Models

Epistemic Fragility in Large Language Models: Prompt Framing Systematically Modulates Misinformation Correction

Reading Between the Lines: Abstaining from VLM-Generated OCR Errors via Latent Representation Probes

Minimax Multi-Target Conformal Prediction with Applications to Imaging Inverse Problems

Modeling and Calibration of Supplier Selection Problem in Freight Agent-Based Simulations

From Classification to Cross-Modal Understanding: Leveraging Vision-Language Models for Fine-Grained Renal Pathology

AccKV: Towards Efficient Audio-Video LLMs Inference via Adaptive-Focusing and Cross-Calibration KV Cache Optimization

Preserving Cross-Modal Consistency for CLIP-based Class-Incremental Learning

LISA: A Layer-wise Integration and Suppression Approach for Hallucination Mitigation in Multimodal Large Language Models