Agent Capability Evaluation on ACEBench Agent
95Multi-Step Reasoning ScoreGPT-4.1
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GPT-4.12026.01 | 95 | 60 | 77.5 | |
| Qwen3-235B-Thinking-2507Mode=Thinking2026.01 | 85 | 63.33 | 74.17 | |
| Kimi-K2-Instruct-09052026.01 | 85 | 73.33 | 79.17 | |
| EnvScalerBackbone=Qwen3-8B, Training=SFT & RL2026.01 | 85 | 60 | 72.5 | |
| EnvScalerBackbone=Qwen3-8B, Training=SFT2026.01 | 83.33 | 60 | 71.67 | |
| Qwen3-235B-Instruct-2507Mode=Instruct2026.01 | 80 | 63.33 | 71.67 | |
| EnvScalerBackbone=Qwen3-4B, Training=SFT & RL2026.01 | 80 | 61.11 | 70.55 | |
| EnvScalerBackbone=Qwen3-4B, Training=SFT2026.01 | 73.33 | 60 | 66.67 | |
| EnvScalerBackbone=Qwen3-1.7B, Training=SFT & RL2026.01 | 66.67 | 33.33 | 50 | |
| Qwen3-8BMode=Thinking2026.01 | 63.33 | 56.67 | 60 | |
| Qwen3-4BMode=Thinking2026.01 | 58.33 | 52.22 | 55.28 | |
| EnvScalerBackbone=Qwen3-1.7B, Training=SFT2026.01 | 55 | 32.22 | 43.61 | |
| Qwen3-1.7BMode=Thinking2026.01 | 35 | 28.89 | 31.95 |