Multiple-Choice Question Answering on PhyX 3k (test)
84.2Exact Match AccuracyQwen3-VL-32B-Thinking
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen3-VL-32B-Thinking2026.05 | 84.2 | |
| Claude Sonnet 4.5max_tokens=163842026.05 | 80.6 | |
| Physics-R1 (dense)max_tokens=16384, reward_type=dense2026.05 | 77.5 | |
| Physics-R1 (binary, seed 42)max_tokens=16384, reward_type=binary, seed=422026.05 | 76.9 | |
| Physics-R1 (binary, 3-seed mean ±σ)max_tokens=16384, reward_type=binary, aggregation=3-seed mean2026.05 | 76.9 | |
| Qwen3-VL-8B-Thinking (base)2026.05 | 74.4 | |
| GPT-4osource=Shen et al. [2025]2026.05 | 53.6 | |
| Gemini 2.5 Promeasured_by=authors2026.05 | 49.8 | |
| InternVL3-8B2026.05 | 43.1 |