Multi-step tool-use composition on DepthBench (val)
98.57Overall AccuracyTIER
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| TIERReward Model=TIER, Method=RL2026.05 | 98.57 | 98.04 | 99.38 | 97.89 | 100 | 93.33 | 90 | 90 | |
| GPT-5Reward Model=GPT-5, Method=3-shot ICL2026.05 | 92.16 | 98.02 | 90.11 | 96.36 | 97.28 | 87.59 | 95 | 60 | |
| Qwen3-8BReward Model=Qwen3-8B, Method=3-shot ICL2026.05 | 79.56 | 88.24 | 85.25 | 72.12 | 59.55 | 69.57 | 57.5 | 50.25 | |
| ToolRLReward Model=ToolRL, Method=RL2026.05 | 67.49 | 92.16 | 90 | 14.74 | 17.33 | 24.44 | 0 | 0 | |
| SimpleReward Model=Simple, Method=RL2026.05 | 66.15 | 98.04 | 94.38 | 1.25 | 3.33 | 0 | 0 | 0 | |
| Qwen3-8BReward Model=Qwen3-8B, Method=Zero-shot2026.05 | 35.51 | 76.47 | 48.54 | 0 | 0 | 0 | 0 | 0 |