Multi-turn tool-calling on ACEBench
15End-to-End Success RateGPT-5
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GPT-5Generation type=N/A2026.06 | 15 | 77.8 | |
| Qwen3-4B-Think + LoRA (3689)Generation type=offline generation without tool feedback (×)2026.06 | 11 | 70.3 | |
| Qwen3-4B-ThinkingGeneration type=N/A2026.06 | 9 | 69.4 | |
| Qwen3-4B-Think + Full (1000)Generation type=connected generation with tool results (✓)2026.06 | 9 | 69.5 | |
| Qwen3-4B-Think + Full (3689)Generation type=offline generation without tool feedback (×)2026.06 | 8 | 61.2 | |
| Qwen3-4B-InstructGeneration type=N/A2026.06 | 4 | 36.2 |