cs.CL（2026-01-07）

📊 共 48 篇论文 | 🔗 7 篇有代码

🎯 兴趣领域导航

支柱九：具身大模型 (Embodied Foundation Models) (42 🔗6) 支柱二：RL算法与架构 (RL & Architecture) (6 🔗1)

🔬 支柱九：具身大模型 (Embodied Foundation Models) (42 篇)

#题目一句话要点标签🔗⭐
1 Layer-Order Inversion: Rethinking Latent Multi-Hop Reasoning in Large Language Models 揭示大语言模型多跳推理中层序反转现象，提出概率式召回-抽取框架 large language model chain-of-thought ✅
2 Mem-Gallery: Benchmarking Multimodal Long-Term Conversational Memory for MLLM Agents 提出Mem-Gallery基准，评估多模态LLM Agent的长期对话记忆能力 large language model multimodal
3 When Helpers Become Hazards: A Benchmark for Analyzing Multimodal LLM-Powered Safety in Daily Life 提出SaLAD：用于评估多模态LLM在日常生活中安全性的基准 large language model multimodal ✅
4 DiffCoT: Diffusion-styled Chain-of-Thought Reasoning in LLMs DiffCoT：利用扩散模型改进LLM中的思维链推理，提升鲁棒性和纠错能力 large language model chain-of-thought
5 ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models ContextFocus：一种激活引导方法，提升大语言模型中的上下文忠实度 large language model
6 Value-Action Alignment in Large Language Models under Privacy-Prosocial Conflict 提出基于上下文的评估协议，分析LLM在隐私与亲社会冲突下的价值-行为一致性 large language model
7 Bridging the Discrete-Continuous Gap: Unified Multimodal Generation via Coupled Manifold Discrete Absorbing Diffusion 提出CoM-DAD，通过耦合流形离散吸收扩散实现统一多模态生成。 multimodal
8 RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models RedBench：构建通用数据集，全面评估大型语言模型的对抗鲁棒性 large language model ✅
9 PsychEthicsBench: Evaluating Large Language Models Against Australian Mental Health Ethics PsychEthicsBench：评估大型语言模型在澳大利亚心理健康伦理方面的表现 large language model
10 How Do Large Language Models Learn Concepts During Continual Pre-Training? 研究大型语言模型概念学习与遗忘机制 large language model
11 From Implicit to Explicit: Token-Efficient Logical Supervision for Mathematical Reasoning in LLMs 提出FSLR框架，通过显式逻辑监督提升LLM在数学推理中的token效率。 large language model chain-of-thought
12 Towards Compositional Generalization of LLMs via Skill Taxonomy Guided Data Synthesis 提出STEPS框架，通过技能分类引导的数据合成提升LLM的组合泛化能力 large language model instruction following
13 From Chains to Graphs: Self-Structured Reasoning for General-Domain LLMs 提出自结构图推理(SGR)框架，提升LLM在开放域问答中的推理一致性 large language model chain-of-thought
14 HearSay Benchmark: Do Audio LLMs Leak What They Hear? HearSay基准测试揭示音频大语言模型存在严重语音隐私泄露风险 large language model chain-of-thought ✅
15 Do LLM Self-Explanations Help Users Predict Model Behavior? Evaluating Counterfactual Simulatability with Pragmatic Perturbations 评估LLM自解释在反事实模拟中的作用：基于语用扰动的用户行为预测 large language model chain-of-thought
16 Persona-aware and Explainable Bikeability Assessment: A Vision-Language Model Approach 提出Persona感知的可解释自行车友好度评估视觉-语言模型 chain-of-thought
17 NeuronScope: A Multi-Agent Framework for Explaining Polysemantic Neurons in Language Models NeuronScope：一种用于解释语言模型中多义神经元的多智能体框架 large language model
18 AI Generated Text Detection 评估AI文本检测方法，提出基于主题划分的基准测试，提升模型泛化能力。 large language model
19 Do LLMs Really Memorize Personally Identifiable Information? Revisiting PII Leakage with a Cue-Controlled Memorization Framework 提出Cue-Resistant Memorization框架以评估LLMs的PII泄露问题 large language model
20 Membox: Weaving Topic Continuity into Long-Range Memory for LLM Agents Membox：通过编织主题连续性增强LLM Agent的长期记忆 large language model
21 Evaluation of Multilingual LLMs Personalized Text Generation Capabilities Targeting Groups and Social-Media Platforms 评估多语言LLM针对群体和社交媒体平台的个性化文本生成能力 large language model
22 ADEPT: Adaptive Dynamic Early-Exit Process for Transformers ADEPT：Transformer的自适应动态早退机制，提升推理效率。 large language model
23 Evaluating LLMs for Police Decision-Making: A Framework Based on Police Action Scenarios 提出PAS框架，评估LLM在警务决策中的应用，解决现有评估体系缺失问题 large language model
24 Reasoning Pattern Alignment Merging for Adaptive Reasoning 提出RPAM：一种基于特征对齐的模型融合框架，用于自适应推理。 chain-of-thought
25 Beyond Perplexity: A Lightweight Benchmark for Knowledge Retention in Supervised Fine-Tuning 提出KR-Test，用于评估SFT中LLM的知识保留能力，区分事实学习与语言模仿。 large language model
26 LLMberjack: Guided Trimming of Debate Trees for Multi-Party Conversation Creation LLMberjack：引导式辩论树修剪平台，用于创建多方对话 large language model
27 Modular Prompt Optimization: Optimizing Structured Prompts with Section-Local Textual Gradients 模块化Prompt优化（MPO）：利用分段局部文本梯度优化结构化Prompt large language model
28 Simulated Students in Tutoring Dialogues: Substance or Illusion? 提出学生模拟任务评估框架，揭示LLM在辅导对话中模拟学生的局限性 large language model
29 Benchmark^2: Systematic Evaluation of LLM Benchmarks Benchmark^2框架：系统评估LLM基准测试质量，提升评估效率。 large language model
30 SoK: Privacy Risks and Mitigations in Retrieval-Augmented Generation Systems 系统性分析RAG系统中隐私风险与缓解措施，构建风险分类与流程图谱。 large language model
31 Decide Then Retrieve: A Training-Free Framework with Uncertainty-Guided Triggering and Dual-Path Retrieval 提出DTR框架，通过不确定性引导触发和双路检索，提升开放域问答中检索增强生成效果。 large language model ✅
32 Evaluating Small Decoder-Only Language Models for Grammar Correction and Text Simplification 评估小型Decoder-Only语言模型在语法纠错和文本简化任务中的性能 large language model
33 Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning 提出ATLAS以解决异构模型与工具的动态选择问题 large language model
34 PartisanLens: A Multilingual Dataset of Hyperpartisan and Conspiratorial Immigration Narratives in European Media PartisanLens：构建欧洲媒体中多语种极端党派和阴谋论移民叙事数据集。 large language model
35 Whose Facts Win? LLM Source Preferences under Knowledge Conflicts 研究知识冲突下LLM对来源偏好，提出方法缓解重复偏见并保持偏好一致性。 large language model
36 SyncThink: A Training-Free Strategy to Align Inference Termination with Reasoning Saturation SyncThink：一种免训练策略，使推理终止与推理饱和度对齐，降低CoT开销。 chain-of-thought
37 ELO: Efficient Layer-Specific Optimization for Continual Pretraining of Multilingual LLMs ELO：面向多语言LLM持续预训练的高效层特异性优化方法 large language model
38 Agent-Dice: Disentangling Knowledge Updates via Geometric Consensus for Agent Continual Learning Agent-Dice：通过几何共识解耦知识更新，解决Agent持续学习中的灾难性遗忘问题 large language model
39 Reasoning Model Is Superior LLM-Judge, Yet Suffers from Biases 研究表明：推理模型作为LLM评判者更优，但仍存在偏见 instruction following
40 DiVA: Fine-grained Factuality Verification with Agentic-Discriminative Verifier 提出Agentic Discriminative Verifier (DiVA)用于细粒度的事实性验证。 large language model
41 EvolMem: A Cognitive-Driven Benchmark for Multi-Session Dialogue Memory EvolMem：一个认知驱动的多轮对话记忆评估基准 large language model ✅
42 DeepSynth-Eval: Objectively Evaluating Information Consolidation in Deep Survey Writing DeepSynth-Eval：提出客观评估深度调研写作中信息整合能力的基准 large language model

🔬 支柱二：RL算法与架构 (RL & Architecture) (6 篇)

#题目一句话要点标签🔗⭐
43 MIND: From Passive Mimicry to Active Reasoning through Capability-Aware Multi-Perspective CoT Distillation 提出MIND框架，通过能力感知的多视角CoT蒸馏实现主动推理，提升小模型的泛化能力。 distillation large language model chain-of-thought
44 O-Researcher: An Open Ended Deep Research Model via Multi-Agent Distillation and Agentic RL O-Researcher：通过多智能体蒸馏和Agentic RL实现开放域深度研究模型 reinforcement learning distillation large language model
45 Step Potential Advantage Estimation: Harnessing Intermediate Confidence and Correctness for Efficient Mathematical Reasoning 提出SPAE，利用中间置信度和正确性进行高效数学推理，提升奖励分配。 reinforcement learning large language model chain-of-thought ✅
46 OLA: Output Language Alignment in Code-Switched LLM Interactions OLA：提出用于评估LLM在混合语境中输出语言对齐的基准测试。 DPO large language model chain-of-thought
47 NeoAMT: Neologism-Aware Agentic Machine Translation with Reinforcement Learning 提出NeoAMT：一种基于强化学习和Wiktionary的感知新词的Agentic机器翻译框架 reinforcement learning reward design
48 KDCM: Reducing Hallucination in LLMs through Explicit Reasoning Structures KDCM：通过显式推理结构减少大型语言模型中的幻觉 distillation large language model

⬅️ 返回 cs.CL 首页 · 🏠 返回主页