Research map / Large language models
Reasoning and chain of thought: research map
1,574 papers on Reasoning and chain of thought in Large language models, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 3 clusters and 12 approaches. The busiest year so far is 2026.
What it is
Large language models solve harder problems when they write out intermediate steps ("chain of thought") before answering. This topic studies how to train and prompt for better reasoning, how to spend more computation at test time (longer thinking, search, verifiers), and how to serve these long outputs quickly.
Terms you will see on this page:
- Chain of thought (CoT)
- the step-by-step reasoning a model writes before its final answer.
- Test-time compute
- extra computation at inference, such as sampling many answers or thinking longer.
- Speculative decoding
- a small model drafts tokens and the large model checks them, which speeds up generation.
Start reading with the most cited papers here:
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (NeurIPS 2022 · 2,251 citations)
- Self-Consistency Improves Chain of Thought Reasoning in Language Models (ICLR 2023 · 706 citations)
Within Large language models, its share grew from 6.5% in 2023–24 to 20.6% in 2025–26 (130 → 1,417 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore Reasoning and chain of thought in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- RL with verifiable rewards for reasoning rising: 0% → 19% of the topic (0 → 270 papers)
- Efficient and adaptive thinking rising: 5% → 16% of the topic (6 → 225 papers)
- Test-time compute scaling rising: 3% → 14% of the topic (4 → 200 papers)
- Search and self-correction in reasoning falling: 25% → 6% of the topic (33 → 87 papers)
- Chain-of-thought training and analysis falling: 18% → 7% of the topic (23 → 104 papers)
Papers from 2025–2026 that are already cited most:
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · 29 citations)
- DAPO: An Open-Source LLM Reinforcement Learning System at Scale (NeurIPS 2025 · 19 citations)
- ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning (NeurIPS 2025 · 10 citations)
- T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · 9 citations)
Approaches and key papers
Test-time scaling · 890 papers
Keywords: test time, scaling, rl
Approaches in this cluster:
- Efficient and adaptive thinking (232 papers)
Control reasoning length and self-verification so models think wisely rather than longer. - Test-time compute scaling (207 papers)
Allocate inference-time compute through confidence, branching, and self-consistency strategies. - RL with verifiable rewards for reasoning (275 papers)
Train reasoning LLMs with reinforcement learning and analyze mathematical reasoning gains. - Routing, pruning and distillation for reasoning (137 papers)
Reduce reasoning cost via model routing, small-large collaboration, pruning, and distillation. - Process reward models (39 papers)
Train process-level reward models with synthetic annotation and calibration to verify reasoning steps.
Most cited and most cited since 2024:
- Measuring Mathematical Problem Solving With the MATH Dataset (NeurIPS 2021 · 276 citations)
- Inference-Time Intervention: Eliciting Truthful Answers from a Language Model (NeurIPS 2023 · 56 citations)
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · 29 citations)
- ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search (NeurIPS 2024 · 21 citations)
Chain-of-thought reasoning · 419 papers
Keywords: chain thought, thought cot, cot reasoning
Approaches in this cluster:
- Chain-of-thought training and analysis (130 papers)
Train and analyze chain-of-thought reasoning, including its faithfulness and information content. - Long CoT dynamics and efficiency (138 papers)
Study the geometry, legibility, and efficiency of long reasoning chains. - Search and self-correction in reasoning (123 papers)
Improve reasoning with trial-and-error, calibration, rollback, and selection-inference loops. - Theory of autoregressive reasoning (28 papers)
Analyze compositional generalization and learnability of reasoning by next-token predictors.
Most cited and most cited since 2024:
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (NeurIPS 2022 · 2,251 citations)
- Self-Consistency Improves Chain of Thought Reasoning in Language Models (ICLR 2023 · 706 citations)
- Graph of Thoughts: Solving Elaborate Problems with Large Language Models (AAAI 2024 · 512 citations)
- MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark (NeurIPS 2024 · 82 citations)
Fast decoding and KV cache · 265 papers
Keywords: decoding, speculative, cache
Approaches in this cluster:
- Efficient LLM serving (127 papers)
Schedule queries, reuse prefixes, and optimize architectures for efficient LLM inference. - Speculative decoding (93 papers)
Use draft models and acceptance-rate optimization to decode faster without changing outputs. - KV cache compression (45 papers)
Evict or merge key-value cache entries to cut memory in LLM inference.
Most cited and most cited since 2024:
- ARTrackV2: Prompting Autoregressive Tracker Where to Look and How to Describe (CVPR 2024 · 108 citations)
- H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models (NeurIPS 2023 · 66 citations)
- SnapKV: LLM Knows What You are Looking for Before Generation (NeurIPS 2024 · 51 citations)
- Hot or Cold? Adaptive Temperature Sampling for Code Generation with Large Language Models (AAAI 2024 · 34 citations)
Related topics in Large language models
- Instruction and fine-tuning (888)
- Retrieval and knowledge (1,084)
- Agents, code and math (1,748)
- Alignment and preferences (672)
- AI and society (913)
- Language, safety and interpretability (3,002)
