Research map / Large language models / Agents, code and math
Scientific logicality enriched methodology for LLM reasoning: A practice in physics
Zhaoxin Yu, Nan Xu, Kun Chen, Jiahao Zhao, Lei Wang, Wenji Mao
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 |
|---|---|---|---|
| add Causal Connection (O) to Logic-Distill sampling (ablation) | Logic-Distill on R1-Distill-Qwen-7B · PhysLogic | answer score: 38.68 -> 49.73 | Improved |
| add Logical Fidelity (F) to Logic-Distill sampling (ablation) | Logic-Distill on Qwen2.5-7B-Instruct · PhysLogic | answer score: 40.59 -> 44.02 | Improved |
| add Inferential Progress (P) to Logic-Distill sampling (ablation) | Logic-Distill on Qwen2.5-7B-Instruct · PhysLogic | answer score: 38.72 -> 44.02 | Improved |
| logicality-based sampling instead of random sampling (ablation) | Logic-Distill on Qwen2.5-7B-Instruct · PhysLogic | answer score: 28.77 -> 44.02 | Improved |
| add Causal Connection (O) to Logic-Distill sampling (ablation) | Logic-Distill on Llama-3.1-8B · PhysLogic | logicality: 44.05 -> 45.50 | Improved |
| logicality-based sampling instead of random sampling (ablation) | Logic-Distill on R1-Distill-Qwen-7B · PhysLogic | logicality: 40.24 -> 44.14 | Improved |
| fine-tune on RST-sampled data instead of MegaScience (second-best) | Llama-3.1-8B-base · PhysLogic | final-answer accuracy: 31.02 -> 44.67 | Improved |
| fine-tune on RST-sampled data instead of SCP-116k (second-best) | DeepSeek-R1-Distill-Qwen-7B · PhysLogic | final-answer accuracy: 46.30 -> 47.45 | Improved |
| Logic-Distill with half the data (40k) vs strongest baseline | Qwen2.5-7B-Instruct · public physics benchmarks | average accuracy: 34.72 -> 45.04 | Improved |
| Logic-Distill (40k) vs Direct-Distill (80k) | Llama-3.1-8B · public physics benchmarks | average accuracy: 37.98 -> 35.14 | Hurt |
| Logic-Distill vs Direct-Distill at equal 40k data | Llama-3.1-8B-base · public physics benchmarks | average accuracy: 29.38 -> 35.14 | Improved |
| Logic-Distill instead of MegaScience data | Qwen-2.5-14B-Instruct · PhysLogic | answer score: 38.41 -> 45.28 | Improved |
Closest papers by meaning
- UGPhysics: A Comprehensive Benchmark for Undergraduate Physics Reasoning with Large Language Models (ICML 2025 · similarity 0.88)
- SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models (ICML 2024 · similarity 0.88)
- Scaling Physical Reasoning with the PHYSICS Dataset (NeurIPS 2025 · similarity 0.88)
- SciReason: A Controllable Benchmark for Scientific Reasoning in LLMs (NeurIPS 2026 · similarity 0.88)
- ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning (ICML 2025 · similarity 0.88)
- LLM-SRBench: A New Benchmark for Scientific Equation Discovery with Large Language Models (ICML 2025 · similarity 0.88)
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.
