AAAI

Breaking the Safety-Capability Tradeoff: Reinforcement Learning with Verifiable Rewards Maintains Safety Guardrails in LLMs

打破安全性与能力的权衡：具有可验证奖励的强化学习在LLMs中维持安全护栏

Dongkyu Derek Cho, Huan Song, Arijit Ghosh Chowdhury, Haotian An, Yawei Wang, Rohit Thekkanal, Negin Sokhandan, Sharlina Keshava, Hannah Marlowe

AI总结 本文提出通过可验证奖励的强化学习方法，在提升LLM推理能力的同时维持安全性，挑战了传统安全性与能力权衡的假设。

Comments AAAI-26 Workshop on Post-AI Formal Methods

LungNoduleAgent: A Collaborative Multi-Agent System for Precision Diagnosis of Lung Nodules

LungNoduleAgent: 一种用于肺结节精准诊断的协作多智能体系统

Cheng Yang, Hui Jin, Xinlei Yu, Zhipeng Wang, Yaoqun Liu, Fenglei Fan, Dajiang Lei, Gangyong Jia, Changmiao Wang, Ruiquan Ge

AI总结 LungNoduleAgent通过协作多智能体系统提升肺结节诊断的精度与效率

Comments Accepted by AAAI 2026

LOOM: Personalized Learning Informed by Daily LLM Conversations Toward Long-Term Mastery via a Dynamic Learner Memory Graph

LOOM：通过动态学习者记忆图实现每日LLM对话指导的个性化学习以实现长期掌握

Justin Cui, Kevin Pu, Tovi Grossman

AI总结 LOOM通过动态学习者记忆图分析每日LLM对话，生成个性化学习材料以实现长期掌握，兼顾连续性和主动性，提升学习者持续进步与即时相关性的平衡。

Comments Accepted to the PerFM Workshop at AAAI 2026

Knowledge Completes the Vision: A Multimodal Entity-aware Retrieval-Augmented Generation Framework for News Image Captioning

知识完善视觉：一种多模态实体感知检索增强生成框架用于新闻图像描述

Xiaoxing You, Qiang Huang, Lingyu Li, Chi Zhang, Xiaopeng Liu, Min Zhang, Jun Yu

AI总结 MERGE提出了一种多模态实体感知检索增强生成框架，通过构建实体中心知识库和改进跨模态对齐，提升新闻图像描述质量和命名实体识别性能。

Comments Accepted to AAAI 2026

$A^2Flow:$ Automating Agentic Workflow Generation via Self-Adaptive Abstraction Operators

$A^2Flow$：通过自适应抽象运算符自动化代理工作流生成

Mingming Zhao, Xiaokang Wei, Yuanqi Shao, Kaiwen Zhou, Lin Yang, Siwei Rao, Junhui Zhan, Zhitang Chen

AI总结 $A^2Flow$通过自适应抽象运算符实现代理工作流的自动化生成，提升性能并降低资源消耗。

Comments Accepted by AAAI-2026

Democratizing LLM Efficiency: From Hyperscale Optimizations to Universal Deployability

让大语言模型效率普及：从超大规模优化到通用部署

Hen-Hsen Huang

AI总结 本文提出通过改进LLM的效率方法，使其在有限资源和专业知识下仍能有效运作，以实现更广泛的部署和公平性。

Comments 8 pages, accepted as a Blue Sky Talk in AAAI 2026

Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning

连接点：通过代理推理实现无需训练的视觉接地

Liqin Luo, Guangyao Chen, Xiawu Zheng, Yongxing Dai, Yixiong Zou, Yonghong Tian

AI总结 GroundingAgent通过代理推理实现无需微调的视觉接地，达到65.1%的零样本准确率，并在选择阶段实现90%的准确率，展示了LLM推理能力的重要性。

Comments AAAI 2026

PrefixGPT: Prefix Adder Optimization by a Generative Pre-trained Transformer

PrefixGPT：通过生成式预训练变换器优化前缀加法器

Ruogu Ding, Xin Ning, Ulf Schlichtmann, Weikang Qian

AI总结 PrefixGPT通过生成式预训练变换器优化前缀加法器，实现7.7%的ADP改进和79.1%的平均ADP降低。

Comments This is an extended version of the paper accepted by the AAAI-2026 Conference

Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction

Mem-PAL: 向基于记忆的个性化对话助手迈进：长期用户-代理交互

Zhaopei Huang, Qifeng Dai, Guozheng Wu, Xiaopeng Wu, Kehan Chen, Chuan Yu, Xubin Li, Tiezheng Ge, Wenxuan Wang, Qin Jin

AI总结 Mem-PAL提出H$^2$Memory框架，通过合成数据和实验验证，提升长期用户-代理交互的个性化能力。

Comments Accepted by AAAI 2026 (Oral)

Wavelet Enhanced Adaptive Frequency Filter for Sequential Recommendation

小波增强的自适应频率滤波器用于序列推荐

Huayang Xu, Huanhuan Yuan, Guanfeng Liu, Junhua Fang, Lei Zhao, Pengpeng Zhao

AI总结 本文提出小波增强的自适应频率滤波器，通过动态频域滤波和小波特征增强，提升序列推荐的性能和效率。

Comments Accepted by AAAI 2026

TinyChemVL: Advancing Chemical Vision-Language Models via Efficient Visual Token Reduction and Complex Reaction Tasks

TinyChemVL：通过高效视觉标记减少和复杂反应任务推进化学视觉-语言模型

Xuanle Zhao, Shuxin Zeng, Xinyuan Cai, Xiang Cheng, Duzhen Zhang, Xiuyi Chen, Bo Xu

AI总结 TinyChemVL通过高效视觉标记减少和复杂反应任务提升化学视觉-语言模型的效率和推理能力，实现更高效的化学任务处理。

Comments Accepted by AAAI 2026

EvoEmpirBench: Dynamic Spatial Reasoning with Agent-ExpVer

EvoEmpirBench: 基于智能体经验的动态空间推理

Pukun Zhao, Longxiang Wang, Miaowei Wang, Chen Chen, Fanqing Zhou, Haojian Huang

机构 * Guangdong University of Finance and Economics（广东金融学院） ； Chongqing University（重庆大学） ； University of Edinburgh（爱丁堡大学） ； The University of Hong Kong（香港大学） 纠错

AI总结 EvoEmpirBench通过动态空间基准评估模型在部分可观测和动态环境下的空间推理与记忆能力，揭示主流模型的限制并提供未来研究平台。

Comments Accepted by AAAI 2026, 29 pages, 3 figures, 7 tables

Where to Start Alignment? Diffusion Large Language Model May Demand a Distinct Position

对齐何处开始？扩散大语言模型可能需要不同的位置

Zhixin Xie, Xurui Song, Jun Luo

AI总结 本文提出MOSA方法，通过强化学习对齐dLLM中间生成与安全拒绝，提升安全性。

Comments Accepted for oral presentation at AAAI 2026

CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysis

CAMERA：通过微专家冗余分析实现多矩阵联合压缩以优化MoE模型

Yuzhuang Xu, Xu Han, Yuanchi Zhang, Yixuan Wang, Yijun Liu, Shiyu Ji, Qingfu Zhu, Wanxiang Che

AI总结 CAMERA通过微专家冗余分析实现多矩阵联合压缩，提升MoE模型的压缩效率与性能。

Comments Accepted in AAAI 2026

Hashed Watermark as a Filter: Defeating Forging and Overwriting Attacks in Weight-based Neural Network Watermarking

哈希水印作为过滤器：在基于权重的神经网络水印中对抗伪造和覆盖攻击

Yuan Yao, Jin Song, Jian Jin

AI总结 NeuralMark通过哈希水印过滤器和平均池化技术，提升基于权重的神经网络水印鲁棒性，有效防御伪造和覆盖攻击。

Comments Accepted by AAAI 2026

Earth-Adapter: Bridge the Geospatial Domain Gaps with Mixture of Frequency Adaptation

Earth-Adapter: 通过频率适应混合填补地理域的空白

Xiaoxing Hu, Ziyang Gong, Yupei Wang, Yuru Jia, Fei Lin, Dexiang Gao, Ke An, Jianhong Han, Zhuoran Sun, Gen Luo, Gen Luo, Xue Yang

AI总结 Earth-Adapter通过频率适应混合方法有效解决遥感图像中的伪影问题，提升基础模型在遥感任务中的性能。

Comments AAAI 2026 camera ready

From Passive Perception to Active Memory: A Weakly Supervised Image Manipulation Localization Framework Driven by Coarse-Grained Annotations

从被动感知到主动记忆：一种由粗粒注释驱动的弱监督图像篡改定位框架

Zhiqing Guo, Dongdong Xi, Songlin Li, Gaobo Yang

AI总结 BoxPromptIML通过粗粒注释策略和双指导特征融合，实现低注释成本下的高精度图像篡改定位。

Comments Accepted by AAAI 2026

The Curious Case of Analogies: Investigating Analogical Reasoning in Large Language Models

类比的奇特案例：在大型语言模型中探讨类比推理

Taewhoo Lee, Minju Song, Chanwoong Yoon, Jungwoo Park, Jaewoo Kang

AI总结 本研究探讨了大型语言模型在类比推理中的能力，发现其在编码和应用高层关系概念方面表现出有限但新兴的能力，揭示了与人类认知的相似性和差距。

Comments AAAI 2026

Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios

扩展大语言模型推测解码：在大批次场景中的非自回归预测

Luohe Shi, Zuchao Li, Lefei Zhang, Baoyuan Qi, Guoming Liu, Hai Zhao

AI总结 SpecFormer通过整合单向和双向注意力机制，实现大批次场景下的高效LLM推理，降低训练和计算成本。

Comments accepted by AAAI-2026

DAPointMamba: Domain Adaptive Point Mamba for Point Cloud Completion

DAPointMamba：用于点云补全的领域自适应点Mamba

Yinghui Li, Qianyu Zhou, Di Shao, Hao Yang, Ye Zhu, Richard Dazeley, Xuequan Lu

AI总结 DAPointMamba通过跨域补丁级扫描、空间SSM对齐和通道SSM对齐模块，有效解决领域自适应点云补全中的几何和语义差异问题，实现高效且准确的点云补全。

Comments Accepted to AAAI 2026

Text-guided Controllable Diffusion for Realistic Camouflage Images Generation

基于文本引导的可控扩散生成逼真伪装图像

Yuhang Qian, Haiyan Chen, Wentong Li, Ningzhong Liu, Jie Qin

AI总结 本文提出CT-CIG方法，通过文本引导和可控扩散生成逼真且逻辑合理的伪装图像，利用VLM和FIRM模块提升伪装图像质量。

Comments Accepted by AAAI 2026

Learning Subgroups with Maximum Treatment Effects without Causal Heuristics

学习最大治疗效应的子群而不使用因果启发式

Lincen Yang, Zhong Li, Matthijs van Leeuwen, Saber Salehkaleybar

AI总结 本文提出了一种基于结构因果模型的方法，无需使用因果启发式，直接学习具有最大治疗效应的子群。

Comments The full version (including the Appendix). Accepted at AAAI 2026

QiMeng-Kernel: Macro-Thinking Micro-Coding Paradigm for LLM-Based High-Performance GPU Kernel Generation

QiMeng-Kernel：基于宏思维微编码的LLM高性能GPU内核生成范式

Xinguo Zhu, Shaohui Peng, Jiaming Guo, Yunji Chen, Qi Guo, Yuanbo Wen, Hang Qin, Ruizhi Chen, Qirui Zhou, Ke Gao, Yanjun Wu, Chen Zhao, Ling Li

AI总结 QiMeng-Kernel通过宏思维微编码范式，结合强化学习和通用LLM，实现高性能GPU内核生成，准确率和效率均优于现有方法。

Comments 9 pages, 2 figures, accepted by AAAI 2026

Learning Procedural-aware Video Representations through State-Grounded Hierarchy Unfolding

通过基于状态的层次解折叠学习过程感知的视频表示

Jinghan Zhao, Yifei Huang, Feng Lu

AI总结 本文提出TSS框架，通过基于状态的层次解折叠学习过程感知的视频表示，提升任务识别和步骤预测性能。

Comments Accepted by AAAI 2026. 15 pages, 12 figures

"Are We Done Yet?": A Vision-Based Judge for Autonomous Task Completion of Computer Use Agents

我们完成了吗？

Marta Sumyk, Oleksandr Kosovan

AI总结 本文提出了一种基于视觉的评估机制，通过视觉-语言模型评估计算机使用代理的任务完成情况，提升了任务完成的准确性和可靠性。

Comments This work has been accepted to appear at the AAAI 2026 Workshop on Trust and Control in Agentic AI (TrustAgent)

HybriDLA: Hybrid Generation for Document Layout Analysis

HybriDLA：文档布局分析的混合生成

Yufan Chen, Omar Moured, Ruiping Liu, Junwei Zheng, Kunyu Peng, Jiaming Zhang, Rainer Stiefelhagen

AI总结 HybriDLA通过结合扩散和自回归解码，实现文档布局分析的高精度区域预测，达到83.5%的mAP性能。

Comments Accepted by AAAI 2026 (Oral). Project page at https://yufanchen96.github.io/projects/HybriDLA

Distilling Cross-Modal Knowledge via Feature Disentanglement

通过特征解耦进行跨模态知识蒸馏

Junhong Liu, Yuan Zhang, Tao Huang, Wenchao Xu, Renyu Yang

AI总结 本文提出频率解耦的跨模态知识蒸馏方法，通过利用频域特征解耦和平衡跨模态知识转移，提升蒸馏效果。

Comments Accepted by AAAI 2026

PaSE: Prototype-aligned Calibration and Shapley-based Equilibrium for Multimodal Sentiment Analysis

PaSE：原型对齐校准与基于Shapley的均衡用于多模态情感分析

Kang He, Boyu Chen, Yuzhe Ding, Fei Li, Chong Teng, Donghong Ji

AI总结 PaSE通过原型对齐校准和基于Shapley的均衡框架，有效缓解多模态情感分析中的模态竞争问题，提升模型性能。

Comments Accepted by AAAI 2026

To Align or Not to Align: Strategic Multimodal Representation Alignment for Optimal Performance

对齐还是不对齐：为最佳性能的战略多模态表示对齐

Wanlong Fang, Tianle Zhang, Alvin Chan

AI总结 本文研究了显式对齐对模型性能的影响，提出可控对比学习模块，发现最佳对齐强度取决于数据冗余量，为优化单模态编码器提供指导。

Comments Accepted by AAAI 2026. This arXiv version includes additional details and extended appendix

Beyond ReAct: A Planner-Centric Framework for Complex Tool-Augmented LLM Reasoning

超越ReAct：一种以规划为中心的框架，用于复杂工具增强的大语言模型推理

Xiaolong Wei, Yuehu Dong, Xingliang Wang, Xingyu Zhang, Zhejun Zhao, Dongdong Shen, Long Xia, Dawei Yin

AI总结 本文提出了一种以规划为中心的框架，通过全局DAG规划和两阶段训练方法，提升复杂工具增强LLM的推理能力。

Comments Accepted by AAAI 2026