AAAI

GOMPSNR: Reflourish the Signal-to-Noise Ratio Metric for Audio Generation Tasks

GOMPSNR：为音频生成任务复兴信号噪声比度量

Lingling Dai, Andong Li, Cheng Chi, Yifan Liang, Xiaodong Li, Chengshi Zheng

AI总结 本文提出GOMPSNR，通过引入相位距离项改进SNR，提升音频生成任务中客观度量的可靠性，并通过两种新型损失函数优化模型性能。

Comments Accepted by AAAI 2026

GeoSteer: Faithful Chain-of-Thought Steering via Latent Manifold Gradients

GeoSteer: 通过潜在流形梯度实现忠实的思维链引导

Kentaro Kazama, Daiki Shirafuji, Tatsuhiko Saito

AI总结 GeoSteer通过学习高质量CoT轨迹的低维流形，利用梯度引导提升LLM中间推理质量，实验显示在GSM8k数据集上准确率和推理质量均显著提升。

Comments The Third workshop of NeusymBridge @AAAI 2026 (Bridging Neurons and Symbols for NLP and Knowledge Graph Reasoning)

OceanSplat: Object-aware Gaussian Splatting with Trinocular View Consistency for Underwater Scene Reconstruction

OceanSplat: 基于三目视图一致性的物体感知高斯点云法用于水下场景重建

Minseong Kweon, Jinsun Park

AI总结 OceanSplat通过三目视图一致性与深度感知调整，实现高保真水下场景重建与恢复。

Comments Accepted to AAAI 2026. Project page: https://oceansplat.github.io

ConSurv: Multimodal Continual Learning for Survival Analysis

ConSurv：面向生存分析的多模态持续学习

Dianzhi Yu, Conghao Xiong, Yankai Chen, Wenqian Cui, Xinni Zhang, Yifei Zhang, Hao Chen, Joseph J. Y. Sung, Irwin King

AI总结 ConSurv是首个用于生存分析的多模态持续学习方法，通过多阶段混合专家和特征受限重放技术，有效解决灾难性遗忘和复杂模态交互问题。

Comments 14 pages, 4 figures. This is the extended version of the paper accepted at AAAI 2026, which includes all technical appendices and additional experimental details

SurgPub-Video: A Comprehensive Surgical Video Dataset for Enhanced Surgical Intelligence in Vision-Language Model

SurgPub-Video: 一个全面的外科视频数据集，用于增强视觉-语言模型中的外科智能

Yaoqian Li, Xikai Yang, Dunyuan Xu, Yang Yu, Litao Zhao, Xiaowei Hu, Jinpeng Li, Pheng-Ann Heng

AI总结 SurgPub-Video数据集和SurgLLaVA-Video模型通过提供高质量外科视频数据和专门的视觉-语言模型，提升了手术场景分析的智能水平。

Journal ref AAAI-2026

Adaptive Riemannian Graph Neural Networks

自适应黎曼图神经网络

Xudong Wang, Chris Ding, Tongxin Li, Jicong Fan

机构 * AAAI-2026 纠错

AI总结 自适应黎曼图神经网络通过学习连续各向异性度量张量场，有效捕捉图数据的复杂几何异质性，提升在同质和异质数据集上的性能。

Comments Accepted in The Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), Main Technical Track

HCF: Hierarchical Cascade Framework for Distributed Multi-Stage Image Compression

HCF：分布式多阶段图像压缩的分层级联框架

Junhao Cai, Taegun An, Chengjun Jin, Sung Il Choi, Juhyun Park, Changhee Joo

AI总结 HCF通过分层级联框架实现分布式多阶段图像压缩的高码率失真性能和高效计算，优于现有方法在PSNR和BD-Rate上的表现。

Comments Accepted at AAAI 2026 as a Conference Paper (Oral Presentation)

The Case for "Thick Evaluations" of Cultural Representation in AI

文化表征在AI中的‘厚评价’案例

Rida Qadri, Mark Diaz, Ding Wang, Michael Madaio

AI总结 本文提出‘厚评估’框架，通过南亚研讨会研究社区对AI生成文化图像的解读方式，以更细致的测量方法评估AI输出中的文化表征。

Comments 10 pages

Journal ref Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(3), 2067-2080 (2025)

Object-Centric Latent Action Learning

基于对象的潜在动作学习

Albina Klepach, Alexander Nikulin, Ilya Zisman, Denis Tarasov, Alexander Derevyagin, Andrei Polubarov, Nikita Lyubaykin, Igor Kiselev, Vladislav Kurenkov

AI总结 本文提出了一种基于对象的潜在动作学习框架，通过自监督对象中心预训练解耦代理与干扰背景动态，提升动作标签的鲁棒性，从而改善模仿学习和代理适应效率。

Comments Accepted by AAAI 2026 (Oral). Source code: https://github.com/dunnolab/object-centric-lapo

Membership Inference Test: Auditing Training Data in Object Classification Models

成员推断测试：在目标分类模型中审计训练数据

Gonzalo Mancera, Daniel DeAlcala, Aythami Morales, Ruben Tolosana, Julian Fierrez

AI总结 本研究提出适用于目标分类模型的MINT架构，通过分析训练数据使用情况，实现70%-80%的识别精度，提升数据透明度。

Comments Deployable AI (DAI 2025) workshop co-located with AAAI-25

RSOD: Reliability-Guided Sonar Image Object Detection with Extremely Limited Labels

RSOD：基于极有限标签的可靠性引导声纳图像目标检测

Chengzhou Li, Ping Guo, Guanchen Meng, Qi Jia, Jinyuan Liu, Zhu Liu, Xiaokang Liu, Yu Liu, Zhongxuan Luo, Xin Fan

AI总结 RSOD通过可靠性引导的教师-学生框架，在极有限标签下实现声纳图像目标检测，利用伪标签策略提升性能，实验表明其在UATD数据集上表现优异。

Comments Accepted by AAAI 2026,9 pages,10 figures

VILTA: A VLM-in-the-Loop Adversary for Enhancing Driving Policy Robustness

VILTA：一种用于增强驾驶策略鲁棒性的视觉语言模型闭环对抗者

Qimao Chen, Fang Li, Shaoqing Xu, Zhiyi Lai, Zixun Xie, Yuechen Luo, Shengyin Jiang, Hanbing Li, Long Chen, Bing Wang, Yi Zhang, Zhi-Xin Yang

AI总结 VILTA通过整合视觉语言模型到闭环训练中，提升自动驾驶策略在长尾事件中的安全性和鲁棒性。

Comments Accepted to AAAI 2026

Life, Machine Learning, and the Search for Habitability: Predicting Biosignature Fluxes for the Habitable Worlds Observatory

生命、机器学习与宜居性探索：为宜居世界观测站预测生物特征信号流量

Mark Moussa, Amber V. Young, Brianna Isola, Vasuda Trehan, Michael D. Himes, Nicholas Wogan, Giada Arney

AI总结 本文提出两种机器学习模型，用于预测系外行星反射光光谱中的生物特征信号流量，以提高宜居世界观测站等任务的观测效率和科学回报。

Comments 8 pages, 4 figures. Submitted and accepted in AAAI-26 (IAAI Emerging Applications track)

Class-Partitioned VQ-VAE and Latent Flow Matching for Point Cloud Scene Generation

类分区的VQ-VAE与潜在流匹配用于点云场景生成

Dasith de Silva Edirimuni, Ajmal Saeed Mian

AI总结 本文提出类分区VQ-VAE与潜在流匹配模型，实现无需外部数据库的点云场景生成，有效减少重建误差。

Comments Accepted to AAAI 2026, Main Technical Track

NADIR: Differential Attention Flow for Non-Autoregressive Transliteration in Indic Languages

NADIR：非自回归翻译中的差分注意力流

Lakshya Tomar, Vinayak Abrol, Puneet Agarwal

AI总结 NADIR通过差分Transformer和专家混合机制，在多语言转写任务中实现高速且高准确性的非自回归翻译系统。

Comments Accepted at the AAAI Conference on Artificial Intelligence (AAAI 2026)

Concepts from Representations: Post-hoc Concept Bottleneck Models via Sparse Decomposition of Visual Representations

表示法中的概念：通过视觉表示的稀疏分解实现事后概念瓶颈模型

Shizhan Gong, Xiaofan Zhang, Qi Dou

AI总结 本文提出PCBM-ReD，通过视觉表示的稀疏分解，实现对预训练模型的可解释性增强，提升图像分类任务的准确性和可解释性。

Comments AAAI 2026

Towards Temporal Fusion Beyond the Field of View for Camera-based Semantic Scene Completion

超越视野范围的基于摄像头的语义场景补全

Jongseong Bae, Junwoo Ha, Jinnyeong Heo, Yeongin Lee, Ha Young Kim

AI总结 本文提出C3DFusion模块，通过融合历史和当前帧的3D特征，提升基于摄像头的语义场景补全效果，显著优于现有方法。

Comments Accepted to AAAI 2026

DeepProofLog: Efficient Proving in Deep Stochastic Logic Programs

DeepProofLog: 在深度随机逻辑程序中实现高效的证明

Ying Jiao, Rodrigo Castellano Ontiveros, Luc De Raedt, Marco Gori, Francesco Giannini, Michelangelo Diligenti, Giuseppe Marra

AI总结 DeepProofLog通过引入深度随机逻辑程序和马尔可夫决策过程的映射，提升了神经符号AI在复杂证明空间和大规模知识库中的可扩展性。

Comments Accepted as an Oral at AAAI 2026

SpikCommander: A High-performance Spiking Transformer with Multi-view Learning for Efficient Speech Command Recognition

SpikCommander: 一种高性能的脉冲变压器与多视图学习相结合的高效语音命令识别方法

Jiaqi Wang, Liutao Yu, Xiongri Shen, Sihang Guo, Chenlin Zhou, Leilei Zhao, Yi Zhong, Zhiguo Zhang, Zhengyu Ma

AI总结 SpikCommander通过多视图学习和脉冲时间感知自注意力模块，实现了高效语音命令识别，优于现有SNN方法。

Comments Accepted by The Fortieth AAAI Conference on Artificial Intelligence (AAAI 2026)

PADiff: Predictive and Adaptive Diffusion Policies for Ad Hoc Teamwork

PADiff: 预测性与自适应扩散策略用于即兴团队合作

Hohei Chan, Xinzhi Zhang, Antao Xiang, Weinan Zhang, Mengchen Zhao

AI总结 PADiff通过整合队友预测信息，提升在非平稳即兴团队合作场景中的预测与适应能力，实现多模态协作模式的多样化。

Comments Accepted by AAAI 2026

ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders

ProtSAE：通过语义引导的稀疏自编码器解构和解释蛋白质语言模型

Xiangyu Liu, Haodi Lei, Yi Liu, Yang Liu, Wei Hu

AI总结 ProtSAE通过语义引导的稀疏自编码器解构蛋白质语言模型，提升其潜在空间中生物相关特征的可解释性与重建保真度。

Comments Accepted in the 39th AAAI Conference on Artificial Intelligence (AAAI 2026)

Answering the Unanswerable Is to Err Knowingly: Analyzing and Mitigating Abstention Failures in Large Reasoning Models

回答不可回答的问题是明知其错：分析和缓解大推理模型中的回避失败

Yi Liu, Xiangyu Liu, Zequn Sun, Wei Hu

AI总结 本文针对大推理模型在面对不可回答问题时的回避失败问题，提出一种轻量级两阶段方法，通过认知监控与推理干预提升回避率并保持推理性能。

Comments Accepted in the 39th AAAI Conference on Artificial Intelligence (AAAI 2026)

STRIDE-QA: Visual Question Answering Dataset for Spatiotemporal Reasoning in Urban Driving Scenes

STRIDE-QA：用于城市驾驶场景时空推理的视觉问答数据集

Keishi Ishihara, Kento Sasaki, Tsubasa Takahashi, Daiki Shiono, Yu Yamaguchi

AI总结 STRIDE-QA通过大规模视觉问答数据集提升自动驾驶中动态交通场景的时空推理能力，显著提升VLMs在空间定位和未来运动预测中的表现。

Comments Accepted to AAAI 2026 (Oral). project page: https://turingmotors.github.io/stride-qa/

DogFit: Domain-guided Fine-tuning for Efficient Transfer Learning of Diffusion Models

DogFit: 域引导微调用于扩散模型高效迁移学习

Yara Bahram, Mohammadhadi Shateri, Eric Granger

AI总结 DogFit通过域感知引导微调提升扩散模型在小目标领域迁移学习的效率与效果，减少计算开销并提高生成质量。

Comments Accepted for poster presentation at AAAI 2026

Hypothesis Generation via LLM-Automated Language Bias for ILP

通过LLM自动语言偏见进行假设生成

Yang Yang, Jiemin Wu, Yutao Yue

AI总结 本文提出通过LLM自动设计语言偏见，结合ILP求解器生成可解释的逻辑规则，提升假设生成的性能和鲁棒性。

Comments accepted by AAAI 2026 Bridge LMReasoning

ParaMETA: Towards Learning Disentangled Paralinguistic Speaking Styles Representations from Speech

ParaMETA: 向学习解耦的语音语义说话风格表示迈进

Haowei Lou, Hye-young Paik, Wen Hu, Lina Yao

AI总结 ParaMETA通过统一框架学习解耦的语音语义说话风格表示，实现多任务处理和生成任务中的细粒度风格控制。

Comments 9 pages, 7 figures, Accepted to AAAI-26 (Main Technical Track)

DeepRAHT: Learning Predictive RAHT for Point Cloud Attribute Compression

DeepRAHT: 基于稀疏张量的端到端预测RAHT点云属性压缩学习

Chunyang Fu, Tai Qin, Shiqi Wang, Zhu Li

AI总结 DeepRAHT通过稀疏张量实现端到端预测RAHT点云属性压缩，提升压缩效率与鲁棒性。

Comments Accepted by AAAI 2026

Speculative Sampling with Reinforcement Learning

推测采样与强化学习

Chenan Wang, Daniel H. Shi, Haipeng Chen

AI总结 Re-SpS通过强化学习优化草稿树超参数，提升大规模语言模型的生成速度，实现高达5.45倍的速度提升。

Comments Accepted to AAAI 2026

Segment and Matte Anything in a Unified Model

统一模型中的分割与遮罩任何事物

Zezhong Fan, Xiaohan Li, Topojoy Biswas, Kaushiki Nag, Kannan Achan

AI总结 本文提出SAMA，一种基于SAM的轻量级模型，实现高质量的交互式图像分割与遮罩，通过多视图定位编码器和定位适配器提升精度，展现广泛的应用能力。

Comments AAAI 2026

Don't Start Over: A Cost-Effective Framework for Migrating Personalized Prompts Between LLMs

不要重头再来：一种成本有效的LLM之间迁移个性化提示的框架

Ziyi Zhao, Chongming Gao, Yang Zhang, Haoyan Liu, Weinan Gan, Huifeng Guo, Yong Liu, Fuli Feng

AI总结 本文提出PUMA框架，通过参数高效适配器和组基用户选择策略，实现LLM间个性化提示的低成本迁移，实验表明其在计算成本上显著优于全量重训练。

Comments Accepted to AAAI 2026 (Oral). 9 pages, 5 figures