cs.CL（2026-05-08）

📊 共 43 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) (27 🔗4) 支柱二：RL算法与架构 (RL & Architecture) (13 🔗3) 支柱一：机器人控制 (Robot Control) (1) 支柱三：空间感知与语义 (Perception & Semantics) (1) 支柱六：视频提取与匹配 (Video Extraction) (1)

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

#题目一句话要点标签🔗⭐
1 Reliable Chain-of-Thought via Prefix Consistency 提出前缀一致性（Prefix Consistency）方法，通过重采样验证提升思维链推理的可靠性 large language model chain-of-thought ✅
2 Post-training makes large language models less human-like 提出Psych-201数据集并揭示后训练过程导致大语言模型行为对齐度下降的现象 large language model
3 PSK@EEUCA 2026: Fine-Tuning Large Language Models with Synthetic Data Augmentation for Multi-Class Toxicity Detection in Gaming Chat 提出基于合成数据增强的Llama 3.1微调策略，以提升游戏聊天场景下的多类毒性检测性能 large language model
4 Understanding Performance Collapse in Layer-Pruned Large Language Models via Decision Representation Transitions 提出基于决策表示转换的分析框架，揭示大语言模型层剪枝导致性能崩溃的内在机制 large language model
5 Why do Large Language Models Fail in Low-resource Translation? Unraveling the Token Dynamics of Large Language Models for Machine Translation 揭示大语言模型在低资源机器翻译中的失效机制：提出Token激活率（TAR）指标 large language model
6 Rethinking Dense Sequential Chains: Reasoning Language Models Can Extract Answers from Sparse, Order-Shuffling Chain-of-Thoughts 挑战思维链范式：推理大模型具备从稀疏且乱序的思维链中提取答案的能力 chain-of-thought
7 NSMQ Riddles: A Benchmark of Scientific and Mathematical Riddles for Quizzing Large Language Models 提出NSMQ Riddles科学与数学谜题基准，用于评估大语言模型在科学推理方面的能力。 large language model
8 WeatherSyn: An Instruction Tuning MLLM For Weather Forecasting Report Generation 提出WeatherSyn模型与数据集，通过指令微调实现气象预报报告的自动化生成 large language model multimodal
9 LaTER: Efficient Test-Time Reasoning via Latent Exploration and Explicit Verification 提出LaTER推理框架：通过潜空间探索与显式验证实现高效测试时推理 large language model chain-of-thought ✅
10 Retrieve, Integrate, and Synthesize: Spatial-Semantic Grounded Latent Visual Reasoning 提出RIS框架：通过空间-语义接地实现多模态大模型的潜在视觉推理 large language model multimodal
11 Beyond Confidence: Rethinking Self-Assessments for Performance Prediction in LLMs 引入认知评价理论：通过多维自我评估提升大语言模型性能预测的可靠性 large language model
12 PaT: Planning-after-Trial for Efficient Test-Time Code Generation 提出PaT（试后规划）框架，通过自适应规划策略显著提升大模型代码生成的推理效率。 large language model
13 Hallucination Detection via Activations of Open-Weight Proxy Analyzers 提出基于开源代理分析器激活值的幻觉检测框架，实现对闭源与开源LLM的通用幻觉识别。 large language model
14 The Memory Curse: How Expanded Recall Erodes Cooperative Intent in LLM Agents 揭示大模型“记忆诅咒”：长上下文窗口如何削弱多智能体协作意图 chain-of-thought
15 Benchmarking EngGPT2-16B-A3B against Comparable Italian and International Open-source LLMs 评估EngGPT2MoE-16B-A3B模型：一种面向意大利语境的高性能混合专家（MoE）大语言模型 large language model
16 Reformulating KV Cache Eviction Problem for Long-Context LLM Inference 提出LaProx框架：通过输出感知矩阵近似重构KV Cache驱逐策略，实现长文本推理的高效压缩 large language model
17 The Text Uncanny Valley: Non-Monotonic Performance Degradation in LLM Information Retrieval 揭示大语言模型中的“文本恐怖谷”现象：非单调性能退化与模式转换机制 large language model
18 Region4Web: Rethinking Observation Space Granularity for Web Agents 提出Region4Web框架：通过功能区域粒度重构网页观测空间，提升Web智能体任务成功率 large language model
19 LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling 提出AutoTTS框架，通过智能体自动发现推理时扩展（TTS）策略以优化计算分配 large language model ✅
20 GLiGuard: Schema-Conditioned Classification for LLM Safeguard 提出GLiGuard：一种基于模式条件分类的轻量级大模型安全防护框架 large language model ✅
21 How Value Induction Reshapes LLM Behaviour 揭示价值诱导对大语言模型行为的影响：安全性、拟人化与谄媚倾向的权衡 large language model
22 Beyond "I cannot fulfill this request": Alleviating Rigid Rejection in LLMs via Label Enhancement 提出LANCE以解决大型语言模型的刚性拒绝问题 large language model
23 Multi-Dimensional Evaluation of LLMs for Grammatical Error Correction 多维度评估大语言模型在语法纠错中的表现：揭示模型性能与评价指标的局限性 large language model
24 Is She Even Relevant? When BERT Ignores Explicit Gender Cues 通过检查点级分析揭示荷兰语BERT模型中性别偏见的形成机制与上下文处理局限 large language model
25 TCMIIES: A Browser-Based LLM-Powered Intelligent Information Extraction System for Academic Literature 提出TCMIIES系统：一种基于浏览器且由LLM驱动的学术文献结构化信息提取平台 large language model
26 From 0-Order Selection to 2-Order Judgment: Combinatorial Hardening Exposes Compositional Failures in Frontier LLMs 提出LogiHard框架：通过组合硬化技术揭示前沿大模型在逻辑推理中的组合性缺陷 zero-shot transfer
27 SAGE: Hierarchical LLM-Based Literary Evaluation through Ontology-Grounded Interpretive Dimensions 提出SAGE分层评估框架，利用本体论驱动的大语言模型实现文学质量的量化评估 large language model

🔬 支柱二：RL算法与架构 (RL & Architecture) (13 篇)

#题目一句话要点标签🔗⭐
28 SEIF: Self-Evolving Reinforcement Learning for Instruction Following 提出SEIF自进化强化学习框架，通过指令难度与模型能力的协同演进提升LLM指令遵循能力。 reinforcement learning large language model instruction following ✅
29 Topology-Enhanced Alignment for Large Language Models: Trajectory Topology Loss and Topological Preference Optimization 提出基于拓扑增强的大语言模型对齐框架，通过轨迹拓扑损失与拓扑偏好优化提升模型生成质量 RLHF DPO large language model
30 Structural Rationale Distillation via Reasoning Space Compression 提出推理路径压缩蒸馏（D-RPC）框架，通过结构化推理路径库提升小模型蒸馏效率与一致性。 distillation large language model chain-of-thought
31 Chain-based Distillation for Effective Initialization of Variable-Sized Small Language Models 提出基于链式蒸馏（CBD）的参数初始化范式，实现变尺度小语言模型的高效训练 distillation large language model
32 SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation 提出SimCT方法，通过跨分词器对齐恢复策略解决异构模型蒸馏中的监督信号丢失问题。 policy learning teacher-student distillation ✅
33 Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs 利用强化学习解锁大语言模型中的参数化知识，而非仅限于推理任务 reinforcement learning chain-of-thought
34 SOD: Step-wise On-policy Distillation for Small Language Model Agents 提出SOD：一种基于步进式在线策略蒸馏的框架，以提升小语言模型智能体的推理能力 reinforcement learning distillation ✅
35 Guidance Is Not a Hyperparameter: Learning Dynamic Control in Diffusion Language Models 提出基于强化学习的动态引导策略，解决扩散语言模型中引导尺度固定导致的控制力与质量权衡问题。 reinforcement learning PPO classifier-free guidance
36 Self-Consolidating Language Models: Continual Knowledge Incorporation from Context 提出自巩固语言模型（SCoL）框架，通过元强化学习实现长上下文知识的持续性参数化更新 reinforcement learning large language model
37 TextLDM: Language Modeling with Continuous Latent Diffusion 提出TextLDM：将视觉扩散模型范式迁移至语言建模，实现统一的生成式架构。 flow matching multimodal
38 Not All Tokens Learn Alike: Attention Entropy Reveals Heterogeneous Signals in RL Reasoning 基于注意力熵的RL推理训练：揭示大模型Token级学习信号的异质性与优化机制 reinforcement learning large language model
39 Intent-Driven Semantic ID Generation for Grounded Conversational News Recommendation 提出意图驱动的语义ID生成框架，解决对话式新闻推荐中的隐式意图与幻觉问题。 distillation chain-of-thought
40 Think-with-Rubrics: From External Evaluator to Internal Reasoning Guidance 提出Think-with-Rubrics范式，将评分准则从外部评估器转化为大模型的内部推理引导机制。 reinforcement learning instruction following

🔬 支柱一：机器人控制 (Robot Control) (1 篇)

#题目一句话要点标签🔗⭐
41 Activation Differences Reveal Backdoors: A Comparison of SAE Architectures 利用差异化稀疏自编码器（Diff-SAE）揭示并隔离语言模型中的后门触发特征 manipulation

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

#题目一句话要点标签🔗⭐
42 The Moltbook Files: A Harmless Slopocalypse or Humanity's Last Experiment 发布Moltbook数据集并评估大规模AI智能体交互对语言模型对齐与安全性的影响 affordance

🔬 支柱六：视频提取与匹配 (Video Extraction) (1 篇)

#题目一句话要点标签🔗⭐
43 CLIPer: Tailoring Diverse User Preference via Classifier-Guided Inference-Time Personalization 提出CLIPer框架，通过分类器引导的推理时个性化实现大语言模型的多维偏好定制 HuMoR

⬅️ 返回 cs.CL 首页 · 🏠 返回主页