Research map / Large language models / Agents, code and math
Sim2Reason: Solving Physics Olympiad via Reinforcement Learning on Physics Simulators
Mihir Prabhudesai, Aryan Satpathy, Yangmin Li, Zheyang Qin, Nikash Bhardwaj, Amir Zadeh, Chuan Li, Katerina Fragkiadaki, Deepak Pathak
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 |
|---|---|---|---|
| train with RL on numeric questions instead of reverse or symbolic (ablation) | Qwen2.5-3B RL · IPhO Mechanics | accuracy: 5.84% (reverse), 7.46% (symbolic) -> 13.15% (numeric) | Improved |
| apply shortcut-solution filtering to generated QA (ablation) | Qwen2.5-3B RL · IPhO Mechanics | accuracy: 7.14% -> 13.15% | Improved |
| use RL instead of SFT on 200k teacher-sampled QA (ablation) | Qwen2.5-32B · IPhO Mechanics | accuracy: 15.9% -> 25.2% | Improved |
| SFT on teacher trajectories (ablation) | Qwen2.5-32B · IPhO Mechanics | accuracy: 19.8% -> 15.9% | Hurt |
| train on simulator data instead of DAPO-17K math data (ablation) | Qwen2.5-3B · IPhO Mechanics | accuracy: 9.98 -> 13.15 | Improved |
| mix DAPO-17K with synthetic data vs synthetic only (ablation) | Qwen2.5-3B · IPhO Mechanics | accuracy: 13.15 -> 10.35 | Hurt |
| RL on synthetic simulator data | Qwen2.5-32B · JEEBench | accuracy: 34.38% -> 52.28% | Improved |
| RL on synthetic simulator data | Qwen2.5-32B · MATH 500 | accuracy: 78.4% -> 82.8% | Improved |
| RL on synthetic simulator data | Qwen3-30B · Synthetic Symbolic | accuracy: 8.8% -> 8.0% | Hurt |
| RL on synthetic simulator data | Qwen2.5-3B · IPhO Mechanics | accuracy: 5.68% -> 13.15% | Improved |
| RL on synthetic simulator data | Qwen2.5-32B · PHYSICS Undergraduate | accuracy: 47.9% -> 48.4% | Improved |
| RL on synthetic simulator data | Qwen2.5-32B · PHYSICS Postgraduate | accuracy: 32.2% -> 37.8% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DAPO-32B — DAPO: An Open-Source LLM Reinforcement Learning System at Scale (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- UGPhysics: A Comprehensive Benchmark for Undergraduate Physics Reasoning with Large Language Models (ICML 2025 · similarity 0.88)
- PhysElite: How Far Are LLMs from Solving Olympiad-Level Physics Problems? (NeurIPS 2026 · similarity 0.87)
- Leveraging Online Olympiad-Level Math Problems for LLMs Training and Contamination-Resistant Evaluation (ICML 2025 · similarity 0.86)
- Scaling Physical Reasoning with the PHYSICS Dataset (NeurIPS 2025 · similarity 0.86)
- SATURN: SAT-based Reinforcement Learning to Unleash LLMs Reasoning (NeurIPS 2025 · similarity 0.86)
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.86)
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.
