Multi-turn function calling on BFCL BASE and LONG CONTEXT multi-turn v3
41.88Avg@4 Success RateHINT-SD-Multi
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HINT-SD-Multibackbone=Qwen3-4B-Instruct-2507, distill_strategy=multiple selected failure-relevant steps, rollouts=42026.05 | 41.88 | 48.75 | |
| HINT-SD-Singlebackbone=Qwen3-4B-Instruct-2507, distill_strategy=first failure-relevant step, rollouts=42026.05 | 36.25 | 43.13 | |
| GRPObackbone=Qwen3-4B-Instruct-2507, rollouts=42026.05 | 31.56 | 41.25 | |
| SDPObackbone=Qwen3-4B-Instruct-2507, rollouts=42026.05 | 30.78 | 40 | |
| SFTmode=supervised fine-tuning, backbone=Qwen3-4B-Instruct-25072026.05 | 28.44 | 38.13 | |
| OpenClaw-RLbackbone=Qwen3-4B-Instruct-2507, rollouts=42026.05 | 28.28 | 45 | |
| Initialmode=zero-shot, backbone=Qwen3-4B-Instruct-25072026.05 | 25.94 | 36.25 |