Research map / Large language models / Reasoning and chain of thought
Efficient Inference for Large Reasoning Models: A Survey
Yue Liu, J. F. Wu, Yufei He, Ruihan Gong, Jun Xia, Liang Li, Hongcheng Gao, Hongyu Chen, Baolong Bi, Jiaheng Zhang, Zhiqi Huang, Bryan Hooi and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| lower TokenSkip compression ratio from 1.0 to 0.5 (ablation) | TokenSkip LLaMA-3.1-8B-Instruct · GSM8K | accuracy: 78.20% -> 86.70% | Improved |
| few-shot prompting instead of zero-shot for CoD (ablation) | CoD GPT-4o · GSM8K | accuracy: 84.40% -> 91.10% | Improved |
| few-shot prompting instead of zero-shot for CoD (ablation) | CoD Claude 3.5 Sonnet · GSM8K | accuracy: 65.50% -> 91.40% | Improved |
| train TALE with SFT instead of prompting (ablation) | TALE LLaMA-3.1-8B-Instruct · GSM8K | accuracy: 84.46% -> 74.11% | Hurt |
| use DPO rather than SFT in TALE (ablation) | TALE LLaMA-3.1-8B-Instruct · GSM8K | accuracy: 74.11% -> 78.41% | Improved |
| use thought-level rather than token-level compression in LightThinker (ablation) | LightThinker R1-Distill-Qwen-7B · GSM8K | accuracy: 87.11% -> 90.14% | Improved |
| add topo-rewarding then hybrid-scaling on top of topo-tuning (ablation) | SOLAR Qwen2-VL-7B · GSM8K | accuracy: 84.00% -> 89.02% | Improved |
| initialize SelfBudgeter from s1k rather than GSM data (ablation) | SelfBudgeter R1-Distill-Qwen-1.5B · GSM8K | accuracy: 76.27% -> 81.50% | Improved |
| use separated reward in ConciseRL (ablation) | ConciseRL R1-Distill-Qwen-1.5B · GSM8K | accuracy: 80.9% -> 72.5% | Hurt |
| raise Constrained-CoT word limit from 15 to 100 (ablation) | Constrained-CoT Falcon-40b · GSM8K | accuracy: 31.5% -> 27.4% | Hurt |
| scale ThinkLess from Qwen2.5-7B to 14B (ablation) | ThinkLess · GSM8K | accuracy: 88.40% -> 92.49% | Improved |
| apply Overclocking to R1-Distill-LLaMA-8B instead of Qwen-32B | Overclocking alpha=100 · GSM8K | accuracy: 85.96% -> 39.87% | Hurt |
Closest papers by meaning
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.94)
- How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.93)
- A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.92)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · similarity 0.92)
- Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression (AAAI 2026 · similarity 0.91)
- SmartThinker: Progressive Chain-of-Thought Length Calibration for Efficient Large Language Model Reasoning (ICML 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
