Research map / Large language models / Agents, code and math
PRISM-Physics: Causal DAG-Based Process Evaluation for Physics Reasoning
Wanjia Zhao, Qinwei Ma, Jingzhe Shi, Shirley Wu, Jiaqi Han, Yijia Xiao, Si-Yuan Chen, Xiao Luo, Ludwig Schmidt, James Zou
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 |
|---|---|---|---|
| use PRISM-DAG causal DAG scoring instead of LLM-as-Judge | evaluation framework · Human alignment (70 problems) | Kendall tau_b: 0.294 -> 0.346 | Improved |
| use PRISM-DAG instead of PSAS-S step-independent scoring | evaluation framework · Human alignment (70 problems) | Kendall tau_b: 0.213 -> 0.346 | Improved |
| raise GPT-5 reasoning effort from low to high | GPT-5 · PRISM-Physics text-only (avg) | step-level accuracy: 46.17 -> 54.13 | Improved |
| raise GPT-5 reasoning effort from low to medium | GPT-5 · PRISM-Physics text-only (avg) | step-level accuracy: 46.17 -> 37.55 | Hurt |
| raise GPT-5-mini reasoning effort from low to high | GPT-5-mini · PRISM-Physics text-only (avg) | step-level accuracy: 37.02 -> 48.78 | Improved |
| raise GPT-5-mini reasoning effort from low to high | GPT-5-mini · PRISM-Physics text-only (avg) | final-answer accuracy: 18.71 -> 26.01 | Improved |
| raise GPT-5-mini reasoning effort from low to medium (latency) | GPT-5-mini · PRISM-Physics | latency: 129.0% higher | Hurt |
| add diagrams (multimodal input) | Gemini-2.5-Pro · PRISM-Physics (avg) | step-level accuracy: 41.19 -> 50.49 | Improved |
| add diagrams (multimodal input) | GPT-5 (medium) · PRISM-Physics (avg) | step-level accuracy: 37.55 -> 54.43 | Improved |
| add diagrams (multimodal input) | Claude-sonnet-4 · PRISM-Physics (avg) | final-answer accuracy: 16.54 -> 19.19 | Improved |
| add diagrams (multimodal input) | smaller or weaker models · PRISM-Physics | accuracy | Mixed |
| use reasoning models instead of chat models | LLMs · PRISM-Physics | accuracy: 41.36 -> 43.39 | Improved |
Closest papers by meaning
- Scaling Physical Reasoning with the PHYSICS Dataset (NeurIPS 2025 · similarity 0.90)
- UGPhysics: A Comprehensive Benchmark for Undergraduate Physics Reasoning with Large Language Models (ICML 2025 · similarity 0.89)
- DAG-Math: Graph-Guided Mathematical Reasoning in LLMs (ICLR 2026 · similarity 0.86)
- PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.86)
- PhysElite: How Far Are LLMs from Solving Olympiad-Level Physics Problems? (NeurIPS 2026 · similarity 0.86)
- PhysGym: Benchmarking LLMs in Interactive Physics Discovery with Controlled Priors (NeurIPS 2025 · similarity 0.85)
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.
