ICML 2025 accepted papers on Large language models
538 of the 3,342 papers accepted at ICML 2025 (16.1%) are on Large language models, compared with 11.9% at ICML 2024. They fall into 7 topics.
See ICML 2025 in the interactive map
Working on something in this area? Describe your idea in scime atlas to see the closest accepted papers by meaning and how crowded the spot has become.
Topics and papers
Language, safety and interpretability · 169 papers
- OR-Bench: An Over-Refusal Benchmark for Large Language Models (9 citations)
- Freeze-Omni: A Smart and Low Latency Speech-to-speech Dialogue Model with Frozen LLM (4 citations)
- NoLiMa: Long-Context Evaluation Beyond Literal Matching (4 citations)
- The Surprising Effectiveness of Test-Time Training for Few-Shot Learning (4 citations)
- Automatically Interpreting Millions of Features in Large Language Models (3 citations)
- Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World (3 citations)
- Correlated Errors in Large Language Models (3 citations)
- Automated Red Teaming with GOAT: the Generative Offensive Agent Tester (2 citations)
- FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems (2 citations)
- Idiosyncrasies in Large Language Models (2 citations)
- Weak-to-Strong Jailbreaking on Large Language Models (2 citations)
- Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive? (2 citations)
and 157 more in the interactive map.
Agents, code and math · 89 papers
- Discovering Symbolic Cognitive Models from Human and Animal Behavior (13 citations)
- Automated Hypothesis Validation with Agentic Sequential Falsifications (7 citations)
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (6 citations)
- SWE-Lancer: Can Frontier LLMs Earn $1 Million from Real-World Freelance Software Engineering? (4 citations)
- RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing (3 citations)
- AlphaVerus: Bootstrapping Formally Verified Code Generation through Self-Improving Translation and Treefinement (2 citations)
- Autoformulation of Mathematical Optimization Models Using LLMs (2 citations)
- Large Language-Geometry Model: When LLM meets Equivariance (2 citations)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (2 citations)
- Mastering Board Games by External and Internal Planning with Language Models (2 citations)
- MedRAX: Medical Reasoning Agent for Chest X-ray (2 citations)
- Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence (2 citations)
and 77 more in the interactive map.
Reasoning and chain of thought · 77 papers
- rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking (8 citations)
- Mind Your Step (by Step): Chain-of-Thought can Reduce Performance on Tasks where Thinking Makes Humans Worse (6 citations)
- Demystifying Long Chain-of-Thought Reasoning (3 citations)
- EPIC: Efficient Position-Independent Caching for Serving Large Language Models (2 citations)
- Puzzle: Distillation-Based NAS for Inference-Optimized LLMs (2 citations)
- Compute or Load KV Cache? Why Not Both? (1 citation)
- Demonstration Selection for In-Context Learning via Reinforcement Learning (1 citation)
- Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning (1 citation)
- Free Process Rewards without Process Labels (1 citation)
- Reward-Guided Speculative Decoding for Efficient LLM Reasoning (1 citation)
- STAIR: Improving Safety Alignment with Introspective Reasoning (1 citation)
- Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (1 citation)
and 65 more in the interactive map.
AI and society · 64 papers
- Position: AI Evaluation Should Learn from How We Test Humans (15 citations)
- AAAR-1.0: Assessing AI’s Potential to Assist Research (3 citations)
- Great Models Think Alike and this Undermines AI Oversight (3 citations)
- Position: Formal Mathematical Reasoning—A New Frontier in AI (3 citations)
- Position: Language model developers should report train-test overlap (3 citations)
- Position: Beyond Assistance – Reimagining LLMs as Ethical and Adaptive Co-Creators in Mental Health Care (2 citations)
- A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment (1 citation)
- Detecting Strategic Deception with Linear Probes (1 citation)
- Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge (1 citation)
- Position: Stop treating `AGI' as the north-star goal of AI research (1 citation)
- Position: When Incentives Backfire, Data Stops Being Human (1 citation)
- ResearchTown: Simulator of Human Research Community (1 citation)
and 52 more in the interactive map.
Alignment and preferences · 58 papers
- From Crowdsourced Data to High-quality Benchmarks: Arena-Hard and Benchbuilder Pipeline (6 citations)
- Aligning with Logic: Measuring, Evaluating and Improving Logical Preference Consistency in Large Language Models (3 citations)
- Thinking LLMs: General Instruction Following with Thought Generation (3 citations)
- AlphaPO: Reward Shape Matters for LLM Alignment (1 citation)
- Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment (1 citation)
- MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (1 citation)
- On Teacher Hacking in Language Model Distillation (1 citation)
- Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples (1 citation)
- Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback (1 citation)
- Accelerating Unbiased LLM Evaluation via Synthetic Feedback
- Aligning LLMs by Predicting Preferences from User Writing Samples
- Aligning Spoken Dialogue Models from User Interactions
and 46 more in the interactive map.
Instruction and fine-tuning · 54 papers
- Quantifying Prediction Consistency Under Fine-tuning Multiplicity in Tabular LLMs (4 citations)
- LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection (2 citations)
- Overtrained Language Models Are Harder to Fine-Tune (2 citations)
- EffiCoder: Enhancing Code Generation in Large Language Models through Efficiency-Aware Fine-tuning (1 citation)
- EVOLvE: Evaluating and Optimizing LLMs For In-Context Exploration (1 citation)
- Fast Exact Unlearning for In-Context Learning Data for LLMs (1 citation)
- Just Enough Shifts: Mitigating Over-Refusal in Aligned Language Models with Targeted Representation Fine-Tuning (1 citation)
- Safe Delta: Consistently Preserving Safety when Fine-Tuning LLMs on Diverse Datasets (1 citation)
- Test-Time Learning for Large Language Models (1 citation)
- Understanding Synthetic Context Extension via Retrieval Heads (1 citation)
- A Memory Efficient Randomized Subspace Optimization Method for Training Large Language Models
- Activation Space Interventions Can Be Transferred Between Large Language Models
and 42 more in the interactive map.
Retrieval and knowledge · 27 papers
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models (8 citations)
- sciLaMA: A Single-Cell Representation Learning Framework to Leverage Prior Knowledge from Large Language Models (3 citations)
- From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (2 citations)
- Continuously Updating Digital Twins using Large Language Models (1 citation)
- GIVE: Structured Reasoning of Large Language Models with Knowledge Graph Inspired Veracity Extrapolation (1 citation)
- KBQA-o1: Agentic Knowledge Base Question Answering with Monte Carlo Tree Search (1 citation)
- LaRA: Benchmarking Retrieval-Augmented Generation and Long-Context LLMs – No Silver Bullet for LC or RAG Routing (1 citation)
- AnyEdit: Edit Any Knowledge Encoded in Language Models
- Are LLMs Prescient? A Continuous Evaluation using Daily News as the Oracle
- Grokking in the Wild: Data Augmentation for Real-World Multi-Hop Reasoning with Transformers
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification
- Is Complex Query Answering Really Complex?
and 15 more in the interactive map.