Research map / Large language models / Reasoning and chain of thought
Theory of autoregressive reasoning
Analyze compositional generalization and learnability of reasoning by next-token predictors.
28 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Reasoning and chain of thought, keyword cluster “chain thought · thought cot · cot reasoning”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
10 papers in 2023–24 and 17 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 7.7% to 1.2%.
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
- 2020 · Hybrid Compositional Reasoning for Reactive Synthesis from Finite-Horizon Specifications (AAAI 2020 · 35 citations)
- 2023 · Language Models can Solve Computer Tasks (NeurIPS 2023 · 67 citations)
- 2024 · Stay on Topic with Classifier-Free Guidance (ICML 2024 · 5 citations)
- 2025 · T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · 9 citations)
- 2026 · Frontier Models Can Take Actions at Low Probabilities (ICML 2026 · 0 citations)
Earliest papers here
- Hybrid Compositional Reasoning for Reactive Synthesis from Finite-Horizon Specifications (AAAI 2020 · 35 citations)
- Language Models can Solve Computer Tasks (NeurIPS 2023 · 67 citations)
- Guiding Large Language Models via Directional Stimulus Prompting (NeurIPS 2023 · 28 citations)
What has been tried
18 papers here report 179 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 6%, hurt 21%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add semantic-level CoT to token-level CoT training | T2I-CompBench · Color | 0.7752 -> 0.8130 | Helped | T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT (NeurIPS 2025 · 9 citations) |
| double in-domain training data | T2I-CompBench · Average score | 50.49 -> 50.67 | Helped | Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO (NeurIPS 2025 · 5 citations) |
| use CFG with γ=1.1 | HumanEval, temperature 0.2 · pass@1 | 11.0% -> 11.8% | Helped | Stay on Topic with Classifier-Free Guidance (ICML 2024 · 5 citations) |
| add supervised fine-tuning after GFlowNet fine-tuning | SUBJ test, 10 training samples · accuracy (%) | 71.4±1.8 -> 75.2±1.8 | Helped | Amortizing intractable inference in large language models (ICLR 2024 · 4 citations) |
| Set rejection probability to 0 instead of the original setting | BBH Logical Deduction · accuracy | 66.3 -> 56.3 | Hurt | Reprompting: Automated Chain-of-Thought Prompt Inference Through Gibbs Sampling (ICML 2024 · 4 citations) |
| Add CoT reasoning to RDES | BANKING77 · accuracy | 0.737 -> 0.838 | Helped | Demonstration Selection for In-Context Learning via Reinforcement Learning (ICML 2025 · 1 citation) |
| Increase the number of probe training datapoints up to 500 | MMLU · consistency and coverage | 500 | Helped | Language Models Can Predict Their Own Behavior (NeurIPS 2025 · 1 citation) |
| Use ICL without CoT | In-distribution parity generalization · accuracy | Hurt | Task Generalization with Autoregressive Compositional Structure: Can Learning from $D$ Tasks Generalize to $D^T$ Tasks? (ICML 2025 · 0 citations) | |
| Replace salient suffix tokens with “and” | Minecraft inference, T=3 and T=5 · Substitution ASR | 0.25; 0.22 | Hurt | Logicbreaks: A Framework for Understanding Subversion of Rule-based Inference (ICLR 2025 · 0 citations) |
| Use hybrid CoT and PoT training instead of CoT-only training | All nine evaluation datasets · average accuracy | 32.0 -> 47.9 | Helped | MAmmoTH: Building Math Generalist Models through Hybrid Instruction Tuning (ICLR 2024) |
What reviewers will expect
Datasets and baselines named in the 17 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GSM-8K · 2 of 17 papers
- RefCOCOg · 2 of 17 papers
- WildJailbreak · 2 of 17 papers
Baselines
- GPT-4o · 3 of 17 papers
- Self-Refine · 2 of 17 papers
- GRPO · 2 of 17 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 27 papers; from OpenAlex and the papers' affiliations).
- Microsoft · 3 papers
- University of Southern California · 2 papers
- Harvard University · 2 papers
- Peking University · 2 papers
- MMLab of CUHK & Shanghai AI Laboratory · 2 papers
- The Chinese University of Hong Kong · 2 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
- Chain-of-thought training and analysis (127 papers)
- Long CoT dynamics and efficiency (136 papers)
- Search and self-correction in reasoning (123 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).