Research map / Large language models / Reasoning and chain of thought
Efficient and adaptive thinking
Control reasoning length and self-verification so models think wisely rather than longer.
231 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · an approach in Reasoning and chain of thought, keyword cluster “test time · scaling · rl”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 4 tracked benchmarks.
6 papers in 2023–24 and 225 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 4.6% to 15.9%.
Open questions the data points to
- GSM8K (accuracy): best went from 85.5 (up to 2024) to 97.73 in 2025 — Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · 0 citations)
- StrategyQA (accuracy): best went from 77.2 (up to 2024) to 85.7 in 2025 — Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · 0 citations)
- Papers report QwQ-32B on OlympiadBench (accuracy (%)) anywhere from 31.25 to 79.7: comparisons against it depend on whose number you take.
- MATH500 (accuracy) is new: first reported in 2025–26 by 13 papers.
- AIME24 (accuracy) is new: first reported in 2025–26 by 6 papers.
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
- 2023 · Self-Evaluation Guided Beam Search for Reasoning (NeurIPS 2023 · 10 citations)
- 2024 · Premise Order Matters in Reasoning with Large Language Models (ICML 2024 · 3 citations)
- 2025 · The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · 29 citations)
- 2026 · MedS³: Towards Medical Slow Thinking with Self-Evolved Soft Dual-sided Process Supervision (AAAI 2026 · 3 citations)
Earliest papers here
- Self-Evaluation Guided Beam Search for Reasoning (NeurIPS 2023 · 10 citations)
- Learning Math Reasoning from Self-Sampled Correct and Partially-Correct Solutions (ICLR 2023 · 2 citations)
- Neural Algorithmic Reasoning Without Intermediate Supervision (NeurIPS 2023 · 1 citation)
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.
MATH500 · accuracy (higher is better)
13 papers, 2025–2026 · first reported in 2025–26
| Year | Best so far | Set by |
|---|---|---|
| 2025 | 96.6 | Teaching Language Models to Reason with Tools (NeurIPS 2025 · 0 citations) |
GSM8K · accuracy (higher is better)
8 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 85.5 | Self-Evaluation Guided Beam Search for Reasoning (NeurIPS 2023 · 10 citations) |
| 2025 | 97.73 | Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens (NeurIPS 2025 · 0 citations) |
AIME24 · accuracy (higher is better)
6 papers, 2025–2026 · first reported in 2025–26
| Year | Best so far | Set by |
|---|---|---|
| 2025 | 86.7 | Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Training (NeurIPS 2025 · 0 citations) |
StrategyQA · accuracy (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 77.2 | Self-Evaluation Guided Beam Search for Reasoning (NeurIPS 2023 · 10 citations) |
| 2025 | 85.7 | Policy Guided Tree Search for Enhanced LLM Reasoning (ICML 2025 · 0 citations) |
What has been tried
56 papers here report 688 measured changes (a component added, removed or swapped, compared with a baseline): helped 65%, no clear effect 6%, hurt 25%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Provide the solution algorithm and ask the model to execute its prescribed steps | Tower of Hanoi and Checker Jumping · performance | No clear effect | The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · 29 citations) | |
| Train on step-by-step verified trajectories rather than random samples | MATH · accuracy | 72.4 -> 78.4 | Helped | rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking (ICML 2025 · 8 citations) |
| Increase test-time recurrence from r=1 to r=32 | GSM8K CoT · flexible/strict match | 0.00/0.00 -> 34.80/42.08 | Helped | Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (NeurIPS 2025 · 5 citations) |
| Remove the direct-reward objective L_DR | RewardBench Reasoning · score | 87.0 -> 77.5 | Hurt | Advancing LLM Reasoning Generalists with Preference Trees (ICLR 2025 · 5 citations) |
| add reasoning generation, compared with removing reasoning | Musical Instruments · H@5 | 0.017567 -> 0.02635 | Helped | R$^2$ec: Towards Large Recommender Models with Reasoning (NeurIPS 2025 · 3 citations) |
| Move relevant rules among the beginning, middle, and end while keeping order and number fixed | Logical reasoning · accuracy | 68.0% / 67.0% / 67.0% (8 rules); 36.5% / 36.5% / 35.0% (12 rules) | No clear effect | Premise Order Matters in Reasoning with Large Language Models (ICML 2024 · 3 citations) |
| balance positive and negative ORM labels per question | GSM8K test set labels · step-level accuracy | 0.74 -> 0.73 | Hurt | GLoRe: When, Where, and How to Improve LLM Reasoning via Global and Local Refinements (ICML 2024 · 3 citations) |
| Train with Stage 2 after Stage 1, before pruning | Five math benchmarks · accuracy (%) | 47.6% -> 51.7% | Helped | Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL (NeurIPS 2025 · 2 citations) |
| Remove the explicit thinking process | Four datasets · MAE | Hurt | Think before Recommendation: Autonomous Reasoning-enhanced Recommender (NeurIPS 2025 · 2 citations) | |
| Use dynamic rather than static node budgets | GSM8K · Accuracy | .779 -> .797 | Helped | LiteSearch: Efficient Tree Search with Dynamic Exploration Budget for Math Reasoning (AAAI 2025 · 1 citation) |
| Add understand question tool | Smolbenchmark · pass@1 accuracy | +26.7% | Helped | Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · 1 citation) |
| Select the shortest correct response rather than a random correct response | GSM8K · accuracy | 95.00 -> 95.82 | Helped | Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning (NeurIPS 2025 · 1 citation) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| QwQ-32B | OlympiadBench | accuracy (%) | 31.25 – 79.7 | 2 papers, e.g. NeurIPS 2025 and NeurIPS 2025 |
| DeepSeek-R1 | AIME24 | accuracy | 73.3 – 79.8 | 2 papers, e.g. NeurIPS 2025 and NeurIPS 2025 |
What reviewers will expect
Datasets and baselines named in the 225 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 · 47 of 225 papers
- MATH · 38 of 225 papers
- MATH500 · 30 of 225 papers
- GPQA-Diamond · 18 of 225 papers
- AIME 2025 · 15 of 225 papers
- AIME 2024 · 14 of 225 papers
- AIME24 · 14 of 225 papers
- AIME25 · 13 of 225 papers
Baselines
- GRPO · 19 of 225 papers
- SFT · 12 of 225 papers
- DEER · 9 of 225 papers
- NoThinking · 9 of 225 papers
- Self-Consistency · 8 of 225 papers
- DeepSeek-R1 · 8 of 225 papers
- MCTS · 6 of 225 papers
- CoT · 6 of 225 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 231 papers; from OpenAlex and the papers' affiliations).
- The Chinese University of Hong Kong · 8 papers
- Carnegie Mellon University · 7 papers
- Microsoft · 6 papers
- Google DeepMind · 6 papers
- Tsinghua University · 6 papers
- University of Science and Technology of China · 6 papers
- Shanghai Jiao Tong University · 6 papers
- Renmin University of China · 6 papers
- Tencent · 5 papers
- Google · 5 papers
Use it
Download BibTeX of 23 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
- Test-time compute scaling (205 papers)
- RL with verifiable rewards for reasoning (270 papers)
- Routing, pruning and distillation for reasoning (135 papers)
- Process reward models (39 papers)
Elsewhere in the topic
- Long CoT dynamics and efficiency (136 papers)
- Chain-of-thought training and analysis (127 papers)
- Efficient LLM serving (126 papers)
- Search and self-correction in reasoning (123 papers)
- Speculative decoding (92 papers)
- KV cache compression (45 papers)
- Theory of autoregressive reasoning (28 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).