Research map / Large language models / Reasoning and chain of thought
Routing, pruning and distillation for reasoning
Reduce reasoning cost via model routing, small-large collaboration, pruning, and distillation.
135 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2024–2026) · an approach in Reasoning and chain of thought, keyword cluster “test time · scaling · rl”
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.
11 papers in 2023–24 and 124 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 8.5% to 8.8%.
Open questions the data points to
- MATH (accuracy (%)): best went from 39.36 (up to 2024) to 87.9 in 2025 — GPO: Learning from Critical Steps to Improve LLM Reasoning (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
- 2024 · AutoMix: Automatically Mixing Language Models (NeurIPS 2024 · 12 citations)
- 2025 · Is There No Such Thing as a Bad Question? H4R: HalluciBot for Ratiocination, Rewriting, Ranking, and Routing (AAAI 2025 · 2 citations)
- 2026 · DeepOR: A Deep Reasoning Foundation Model for Optimization Modeling (AAAI 2026 · 2 citations)
Earliest papers here
- AutoMix: Automatically Mixing Language Models (NeurIPS 2024 · 12 citations)
- Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing (ICLR 2024 · 7 citations)
- Ensemble Learning for Heterogeneous Large Language Models with Deep Parallel Collaboration (NeurIPS 2024 · 7 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.
MATH · accuracy (%) (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 39.36 | AutoPSV: Automated Process-Supervised Verifier (NeurIPS 2024 · 5 citations) |
| 2025 | 87.9 | GPO: Learning from Critical Steps to Improve LLM Reasoning (NeurIPS 2025 · 0 citations) |
What has been tried
56 papers here report 609 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 8%, hurt 17%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Replace the threshold router with the POMDP router | Across the five datasets and three small models · ΔIBC (%) | Mixed | AutoMix: Automatically Mixing Language Models (NeurIPS 2024 · 12 citations) | |
| add data transformation (relaxed threshold t* labels) to probabilistic router | MixInstruct large gap, 40% cost advantage · BART score drop (%) | 13.1 -> 10.3 vs r_prob | Helped | Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing (ICLR 2024 · 7 citations) |
| Use full common-token anchors rather than fewer anchors | MMLU and ARC-C · accuracy | Helped | Ensemble Learning for Heterogeneous Large Language Models with Deep Parallel Collaboration (NeurIPS 2024 · 7 citations) | |
| AutoPSV process labels instead of MCTS process labels (labeled GSM8K/MATH) | GSM8K and MATH · selection accuracy (%) | GSM8K Mistral +1.19, Mixtral -0.24, Qwen +0.66; MATH Mistral -0.06, Mixtral +0.36, Qwen +0.22 | Mixed | AutoPSV: Automated Process-Supervised Verifier (NeurIPS 2024 · 5 citations) |
| Use ghost inner-product computation instead of direct per-sample gradients | OPENWEBTEXT · throughput (training data points per second) | 4.2 -> 71.3 | Helped | GREATS: Online Selection of High-Quality Data for LLM Training in Every Iteration (NeurIPS 2024 · 4 citations) |
| Fit scaling-law coefficients using only runs with no more than 100 tokens per parameter | Extreme token-per-parameter training regime · accuracy of estimated loss improvement from additional training data | Hurt | Beyond Chinchilla-Optimal: Accounting for Inference in Language Model Scaling Laws (ICML 2024 · 3 citations) | |
| adapt per-query latency estimate hourly (historical average) in budget denominator instead of fixed latency | GSM8K, time-varying API latency, $0.6 budget · accuracy (%) | 86.4 vs 76.1 | Helped | Efficient Contextual LLM Cascades through Budget-Constrained Policy Learning (NeurIPS 2024 · 3 citations) |
| add scenario tags to the query | HalluciBot binary test · accuracy | 66.1 -> 69.0 | Helped | Is There No Such Thing as a Bad Question? H4R: HalluciBot for Ratiocination, Rewriting, Ranking, and Routing (AAAI 2025 · 2 citations) |
| Use advantage normalization during RL | Math evaluation (reported generally) · Accuracy | Hurt | Training Language Models to Reason Efficiently (NeurIPS 2025 · 2 citations) | |
| Reduce feedback density from 100% to 20% | ARC Challenge · operating cost (MJ) | 0.83 -> 0.79 | Helped | MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees (NeurIPS 2025 · 2 citations) |
| add debate summarization | GSM · accuracy | 82.20 -> 85.60 | Helped | Multiagent Finetuning: Self Improvement with Diverse Reasoning Chains (ICLR 2025 · 2 citations) |
| remove reflection and improvement annotations from self-correction data | GSM8K · P@1 | -1.29 (82.94 to 81.65) | Hurt | S^3cMath: Spontaneous Step-Level Self-Correction Makes Large Language Models Better Mathematical Reasoners (AAAI 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 124 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 124 papers
- MMLU · 19 of 124 papers
- MATH · 14 of 124 papers
- MATH500 · 11 of 124 papers
- MMLU-Pro · 10 of 124 papers
- HotpotQA · 6 of 124 papers
- MT-Bench · 6 of 124 papers
- GPQA · 6 of 124 papers
Baselines
- self-consistency · 9 of 124 papers
- greedy decoding · 5 of 124 papers
- Best-of-N · 5 of 124 papers
- Self-Refine · 5 of 124 papers
- GPT-4o · 5 of 124 papers
- Gemini 2.5 Pro · 4 of 124 papers
- FrugalGPT · 4 of 124 papers
- Beam Search · 4 of 124 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 135 papers; from OpenAlex and the papers' affiliations).
- Carnegie Mellon University · 7 papers
- Hong Kong University of Science and Technology · 6 papers
- Google · 5 papers
- Meta · 5 papers
- Alibaba Group · 5 papers
- Microsoft · 4 papers
- Nanyang Technological University · 4 papers
- Stanford University · 4 papers
- Apple · 3 papers
- Google DeepMind · 3 papers
Use it
Download BibTeX of 20 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
- Efficient and adaptive thinking (231 papers)
- Test-time compute scaling (205 papers)
- RL with verifiable rewards for reasoning (270 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).