Tool-use Generalization on ACEBench Format and Content OOD v1 (test)
38Adherence (Ad.)Vanilla
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| VanillaBackbone=Qwen2.5-1.5B-Instruct2026.06 | 38 | 20 | 46 | 0 | 26 | |
| OPSBackbone=Qwen2.5-1.5B-Instruct2026.06 | 36 | 24 | 43 | 0 | 25.75 | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct2026.06 | 32 | 24 | 43 | 0 | 24.75 | |
| HBGBackbone=Qwen2.5-1.5B-Instruct2026.06 | 28 | 20 | 42 | 0 | 22.5 | |
| PRSBackbone=Qwen2.5-1.5B-Instruct2026.06 | 16 | 8 | 24 | 0 | 12 | |
| ETSBackbone=Qwen2.5-1.5B-Instruct2026.06 | 10 | 0 | 25 | 0 | 8.75 | |
| SFT_BFCLBackbone=Qwen2.5-1.5B-Instruct2026.06 | 0 | 0 | 0 | 0 | 0 | |
| SFT_BFCL + RLBackbone=Qwen2.5-1.5B-Instruct2026.06 | 0 | 0 | 0 | 0 | 0 |