cs.AI（2026-02-02）

📊 共 40 篇论文 | 🔗 3 篇有代码

🎯 兴趣领域导航

支柱九：具身大模型 (Embodied Foundation Models) (27 🔗2) 支柱二：RL算法与架构 (RL & Architecture) (10 🔗1) 支柱三：空间感知与语义 (Perception & Semantics) (1) 支柱四：生成式动作 (Generative Motion) (1) 支柱八：物理动画 (Physics-based Animation) (1)

🔬 支柱九：具身大模型 (Embodied Foundation Models) (27 篇)

#题目一句话要点标签🔗⭐
1 Thinking with Comics: Enhancing Multimodal Reasoning through Structured Visual Storytelling 提出基于漫画的视觉推理范式，提升多模态时序和因果推理能力 large language model multimodal chain-of-thought
2 Entropy-Guided Data-Efficient Training for Multimodal Reasoning Reward Models 提出熵引导训练(EGT)方法，提升多模态推理奖励模型的训练效率与性能。 large language model multimodal
3 Avenir-Web: Human-Experience-Imitating Multimodal Web Agents with Mixture of Grounding Experts Avenir-Web：基于混合专家和经验模仿的多模态Web Agent，提升复杂Web环境下的任务执行能力 large language model multimodal
4 Hunt Instead of Wait: Evaluating Deep Data Research on Large Language Models 提出DDR-Bench，评估LLM在开放数据分析中的自主探索能力 large language model
5 Evolving from Tool User to Creator via Training-Free Experience Reuse in Multimodal Reasoning 提出UCT框架，通过免训练经验复用，使多模态推理Agent从工具使用者进化为创造者 multimodal
6 Large Language Model and Formal Concept Analysis: a comparative study for Topic Modeling 对比研究大型语言模型与形式概念分析在主题建模中的应用 large language model
7 Optimizing Prompts for Large Language Models: A Causal Approach 提出因果提示优化（CPO）框架，解决大语言模型提示工程中的泛化性和成本问题。 large language model
8 MentisOculi: Revealing the Limits of Reasoning with Mental Imagery MentisOculi：揭示心智图像推理的局限性，评估多模态模型利用视觉信息的能力 large language model multimodal
9 Live-Evo: Online Evolution of Agentic Memory from Continuous Feedback 提出Live-Evo以解决在线记忆演化问题 large language model ✅
10 Interpreting and Controlling LLM Reasoning through Integrated Policy Gradient 提出IPG方法，通过积分策略梯度实现对LLM推理过程的解释与控制 large language model
11 Light Alignment Improves LLM Safety via Model Self-Reflection with a Single Neuron 提出基于单神经元门控机制的轻量级对齐方法，提升LLM安全性。 large language model ✅
12 Geometric Analysis of Token Selection in Multi-Head Attention 提出多头注意力几何分析框架，揭示Token选择机制与头部的专门化行为 large language model
13 RedVisor: Reasoning-Aware Prompt Injection Defense via Zero-Copy KV Cache Reuse RedVisor：通过零拷贝KV缓存复用实现推理感知的提示注入防御 large language model
14 PRISM: Parametrically Refactoring Inference for Speculative Sampling Draft Models PRISM：通过参数化重构推理解耦模型容量与推理成本，加速推测采样 large language model
15 Breaking the Reversal Curse in Autoregressive Language Models via Identity Bridge 提出身份桥接方法以解决自回归语言模型的反转诅咒问题 large language model
16 Structure Enables Effective Self-Localization of Errors in LLMs 提出Thought-ICS框架，通过结构化推理实现LLM的有效误差自定位与修正 chain-of-thought
17 More Than a Quick Glance: Overcoming the Greedy Bias in KV-Cache Compression LASER-KV：通过精确LSH召回克服KV缓存压缩中的贪婪偏差 large language model
18 Reasoning in a Combinatorial and Constrained World: Benchmarking LLMs on Natural-Language Combinatorial Optimization 提出NLCO基准，评估LLM在自然语言组合优化问题中的推理能力 large language model
19 See2Refine: Vision-Language Feedback Improves LLM-Based eHMI Action Designers See2Refine：利用视觉-语言反馈提升LLM驱动的eHMI动作设计 large language model
20 Constrained Process Maps for Multi-Agent Generative AI Workflows 提出多代理生成AI工作流的约束过程图以解决不确定性问题 large language model
21 Do I Really Know? Learning Factual Self-Verification for Hallucination Reduction 提出VeriFY框架，通过自验证学习减少大语言模型的事实性幻觉 large language model
22 Human Society-Inspired Approaches to Agentic AI Security: The 4C Framework 提出4C框架，应对Agentic AI在开放环境中涌现的安全风险 large language model
23 GRAB: An LLM-Inspired Sequence-First Click-Through Rate Prediction Modeling Paradigm GRAB：受LLM启发的序列优先点击率预测建模范式，提升广告收益和点击率。 large language model
24 Meta Engine: A Unified Semantic Query Engine on Heterogeneous LLM-Based Query Systems 提出Meta Engine，统一异构LLM语义查询系统，解决多模态数据查询难题。 large language model
25 Beyond Dense States: Elevating Sparse Transcoders to Active Operators for Latent Reasoning 提出LSTR：提升稀疏转码器为主动算子，用于潜在空间推理 chain-of-thought
26 What LLMs Think When You Don't Tell Them What to Think About? 研究揭示：在无主题引导下，大语言模型展现出显著且系统性的主题偏好 large language model
27 The Strategic Foresight of LLMs: Evidence from a Fully Prospective Venture Tournament 大型语言模型在战略预测中超越人类专家，尤其在众筹项目成功预测方面 large language model

🔬 支柱二：RL算法与架构 (RL & Architecture) (10 篇)

#题目一句话要点标签🔗⭐
28 Position: Explaining Behavioral Shifts in Large Language Models Requires a Comparative Approach 提出Δ-XAI框架，用于解释大语言模型行为转变 reinforcement learning large language model foundation model
29 Rethinking the Role of Entropy in Optimizing Tool-Use Behaviors for Large Language Model Agents 利用熵减优化LLM智能体工具使用行为，提升效率与性能 reward design large language model
30 DomusFM: A Foundation Model for Smart-Home Sensor Data DomusFM：面向智能家居传感器数据的预训练基础模型 contrastive learning foundation model
31 Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models 提出GPS，通过小模型预测提示词难度，高效指导大模型强化学习后训练。 reinforcement learning predictive model large language model
32 FlowSteer: Interactive Agentic Workflow Orchestration via End-to-End Reinforcement Learning FlowSteer：通过端到端强化学习实现交互式Agent工作流编排 reinforcement learning large language model
33 Edit Knowledge, Not Just Facts via Multi-Step Reasoning over Background Stories 提出基于背景故事多步推理的知识编辑方法，提升模型知识整合与泛化能力 distillation large language model
34 ProcMEM: Learning Reusable Procedural Memory from Experience via Non-Parametric PPO for LLM Agents ProcMEM：通过非参数PPO从经验中学习可复用程序记忆，用于LLM智能体 PPO
35 TABX: A High-Throughput Sandbox Battle Simulator for Multi-Agent Reinforcement Learning TABX：用于多智能体强化学习的高吞吐量沙盒战斗模拟器 reinforcement learning
36 MAGIC: A Co-Evolving Attacker-Defender Adversarial Game for Robust LLM Safety MAGIC：一种用于增强LLM安全性的协同进化攻防对抗博弈方法 reinforcement learning large language model ✅
37 Adversarial Reward Auditing for Active Detection and Mitigation of Reward Hacking 提出对抗奖励审计框架，主动检测并缓解奖励模型中的奖励攻击问题。 reinforcement learning RLHF

🔬 支柱三：空间感知与语义 (Perception & Semantics) (1 篇)

#题目一句话要点标签🔗⭐
38 MarkCleaner: High-Fidelity Watermark Removal via Imperceptible Micro-Geometric Perturbation MarkCleaner：通过不可察觉的微几何扰动实现高保真水印去除 gaussian splatting splatting

🔬 支柱四：生成式动作 (Generative Motion) (1 篇)

#题目一句话要点标签🔗⭐
39 Understanding the Reversal Curse Mitigation in Masked Diffusion Models through Attention and Training Dynamics 通过注意力和训练动态理解掩码扩散模型中逆转诅咒的缓解 MDM

🔬 支柱八：物理动画 (Physics-based Animation) (1 篇)

#题目一句话要点标签🔗⭐
40 Synesthesia of Vehicles: Tactile Data Synthesis from Visual Inputs 提出车辆触觉感知框架SoV，通过视觉输入预测车辆行驶过程中的触觉激励，提升自动驾驶安全性。 spatiotemporal

⬅️ 返回 cs.AI 首页 · 🏠 返回主页