Research map / Multi-agent systems and games / Human-AI and goal-conditioned
PHYRE: A New Benchmark for Physical Reasoning
Anton Bakhtin, Laurens van der Maaten, Justin Johnson, Laura Gustafson, Ross Girshick
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 online test-time updates | DQN · PHYRE-B cross-template | AUCCESS: 36.8 -> 56.2 | Improved |
| add online test-time updates | DQN · PHYRE-2B cross-template | AUCCESS: 23.2 -> 39.6 | Improved |
| add online test-time updates | MEM · PHYRE-B cross-template | AUCCESS: 18.5 -> 22.8 | Improved |
| add online test-time updates | MEM · PHYRE-2B cross-template | AUCCESS: 3.7 -> 4.9 | Improved |
| add online updates during testing (ablation) | DQN · PHYRE within-template | AUCCESS | Hurt |
| remove class-balancing of positive and negative training examples (ablation) | DQN · PHYRE, particularly PHYRE-2B | AUCCESS | Hurt |
| fuse action features at the input to conv2 instead of conv5 (ablation) | DQN · PHYRE | AUCCESS | No clear effect |
| fuse action features at every stage from conv2 to conv5 (ablation) | DQN · PHYRE | AUCCESS | No clear effect |
| fuse action features with globally pooled conv5 features (ablation) | DQN · PHYRE | AUCCESS | Hurt |
| increase the action encoder to one hidden layer of 1024 units (ablation) | DQN · PHYRE | AUCCESS | Mixed |
| increase the action encoder to two hidden layers of 1024 units (ablation) | DQN · PHYRE | AUCCESS | Mixed |
| rank more candidate actions (ablation) | RAND, MEM, and DQN · PHYRE | AUCCESS | Mixed |
Closest papers by meaning
- I-PHYRE: Interactive Physical Reasoning (ICLR 2024 · similarity 0.90)
- PhysGym: Benchmarking LLMs in Interactive Physics Discovery with Controlled Priors (NeurIPS 2025 · similarity 0.85)
- PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.84)
- ContPhy: Continuum Physical Concept Learning and Reasoning from Videos (ICML 2024 · similarity 0.84)
- PhysElite: How Far Are LLMs from Solving Olympiad-Level Physics Problems? (NeurIPS 2026 · similarity 0.84)
- Causal-PIK: Causality-based Physical Reasoning with a Physics-Informed Kernel (ICML 2025 · similarity 0.84)
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.
