Research map / Large language models / Reasoning and chain of thought
Long CoT dynamics and efficiency
Study the geometry, legibility, and efficiency of long reasoning chains.
136 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Reasoning and chain of thought, keyword cluster “chain thought · thought cot · cot reasoning”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 1 tracked benchmark.
5 papers in 2023–24 and 127 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.8% to 9.0%.
Open questions the data points to
- GSM8K (accuracy): best went from 89.29 (up to 2024) to 96.7 in 2025 — VeriThinker: Learning to Verify Makes Reasoning Model Efficient (NeurIPS 2025 · 0 citations)
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Ontology Materialization by Abstraction Refinement in Horn SHOIF (AAAI 2017 · 13 citations)
- 2018 · Qualitative Reasoning About Cardinal Directions Using Answer Set Programming (AAAI 2018 · 10 citations)
- 2019 · From Horn-SRIQ to Datalog: A Data-Independent Transformation That Preserves Assertion Entailment (AAAI 2019 · 11 citations)
- 2023 · Self-Consistency Improves Chain of Thought Reasoning in Language Models (ICLR 2023 · 706 citations)
- 2024 · CRUXEval: A Benchmark for Code Reasoning, Understanding and Execution (ICML 2024 · 5 citations)
- 2025 · Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (NeurIPS 2025 · 4 citations)
- 2026 · Chain-of-Thought Reasoning In The Wild Is Not Always Faithful (ICML 2026 · 2 citations)
Earliest papers here
- Ontology Materialization by Abstraction Refinement in Horn SHOIF (AAAI 2017 · 13 citations)
- Qualitative Reasoning About Cardinal Directions Using Answer Set Programming (AAAI 2018 · 10 citations)
- From Horn-SRIQ to Datalog: A Data-Independent Transformation That Preserves Assertion Entailment (AAAI 2019 · 11 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
GSM8K · accuracy (higher is better)
5 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 78 | Self-Consistency Improves Chain of Thought Reasoning in Language Models (ICLR 2023 · 706 citations) |
| 2024 | 89.29 | Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoning (ICLR 2024 · 3 citations) |
| 2025 | 96.7 | VeriThinker: Learning to Verify Makes Reasoning Model Efficient (NeurIPS 2025 · 0 citations) |
What has been tried
35 papers here report 356 measured changes (a component added, removed or swapped, compared with a baseline): helped 58%, no clear effect 8%, hurt 26%, mixed 8%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add chain-of-thought prompting | CRUXEval · pass@1 | +7.7% input, +18.5% output | Helped | CRUXEval: A Benchmark for Code Reasoning, Understanding and Execution (ICML 2024 · 5 citations) |
| Remove per-token reward/loss averaging | MATH, DeepSeek-R1-Distill-Qwen-1.5B · reasoning chain length | 9.79 -> 128 | Hurt | Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (NeurIPS 2025 · 4 citations) |
| early stopping with max sample size varied from 16 to 64 (window 8) | MATH · accuracy (%) | 0.00 at every L | No clear effect | Escape Sky-high Cost: Early-stopping Self-Consistency for Multi-step Reasoning (ICLR 2024 · 3 citations) |
| Add Level 2 dropping to Level 1 dropping | Maze 30x30 · 1-Optimal-64 | 97.5 -> 98.1 | Helped | Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces (ICLR 2025 · 2 citations) |
| Use λ=1 instead of λ=10 for k=4 regularization | Reasoning Primitives · accuracy | 47.5 -> 42.5 | Hurt | Reasoning with Latent Thoughts: On the Power of Looped Transformers (ICLR 2025 · 2 citations) |
| Use binary instead of detailed verification in reflective SFT | Mult ID · accuracy | Mixed | Self-Verifying Reflection Helps Transformers with CoT Reasoning (NeurIPS 2025 · 1 citation) | |
| Replace the Safety Primer with a direct caution prefix | MATH500 · accuracy | 84.0 -> 47.6 | Hurt | SAFEPATH: Preventing Harmful Reasoning in Chain-of-Thought via Early Alignment (NeurIPS 2025 · 1 citation) |
| Increase soft-attention scaling factor β | Marginal 3-SAT datasets · SAT/UNSAT accuracy | Helped | Can Transformers Reason Logically? A Study in SAT Solving (ICML 2025 · 0 citations) | |
| Add corrector refinement rounds | MiniF2F-Test · accuracy | 51.64% -> 53.28% -> 54.51% -> 55.33% | Helped | MA-LoT: Model-Collaboration Lean-based Long Chain-of-Thought Reasoning enhances Formal Theorem Proving (ICML 2025 · 0 citations) |
| Provide only every k-th prefix-parity CoT step | Parity · accuracy | Hurt | Lower Bounds for Chain-of-Thought Reasoning in Hard-Attention Transformers (ICML 2025 · 0 citations) | |
| Set α=0, removing the correctness reward while retaining the SOL design | Training · accuracy | Mixed | ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning (NeurIPS 2025 · 0 citations) | |
| use adaptive Q-DYN selection rather than shortest-only selection | AIME24 · accuracy | 53.3 -> 57.3 | Helped | TokenSqueeze: Performance-Preserving Compression for Reasoning LLMs (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 127 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GSM8K · 35 of 127 papers
- MATH500 · 24 of 127 papers
- MATH · 15 of 127 papers
- GPQA-Diamond · 14 of 127 papers
- MMLU · 12 of 127 papers
- AIME 2025 · 12 of 127 papers
- LiveCodeBench · 11 of 127 papers
- AIME 2024 · 9 of 127 papers
Baselines
- DEER · 7 of 127 papers
- DeepSeek-R1 · 7 of 127 papers
- Coconut · 6 of 127 papers
- SEAL · 5 of 127 papers
- Vanilla · 5 of 127 papers
- TokenSkip · 5 of 127 papers
- GPT-4o · 5 of 127 papers
- DeepSeek-V3 · 4 of 127 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 132 papers; from OpenAlex and the papers' affiliations).
- Meta · 4 papers
- Shanghai Jiao Tong University · 4 papers
- Tsinghua University · 4 papers
- Google · 3 papers
- Google DeepMind · 3 papers
- Hong Kong University of Science and Technology · 3 papers
- Harvard University · 3 papers
- Zhejiang University · 3 papers
- National University of Singapore · 3 papers
- Xiaohongshu · 2 papers
Use it
Download BibTeX of 25 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Chain-of-thought training and analysis (127 papers)
- Search and self-correction in reasoning (123 papers)
- Theory of autoregressive reasoning (28 papers)
Elsewhere in the topic
- RL with verifiable rewards for reasoning (270 papers)
- Efficient and adaptive thinking (231 papers)
- Test-time compute scaling (205 papers)
- Routing, pruning and distillation for reasoning (135 papers)
- Efficient LLM serving (126 papers)
- Speculative decoding (92 papers)
- KV cache compression (45 papers)
- Process reward models (39 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).