Multi-turn Tool-use on FTRL, BFCL, and ToolHop Combined
36.9Average Success RatePACT
Evaluation Results
| Method | Links | |
|---|---|---|
| PACTBackbone=Qwen3-4B2026.06 | 36.9 | |
| SFT→MatchTIRBackbone=Qwen3-4B2026.06 | 33.07 | |
| MatchTIRBackbone=Qwen3-4B2026.06 | 32.93 | |
| FTRLBackbone=Qwen3-4B2026.06 | 31.13 | |
| CHORDBackbone=Qwen3-4B2026.06 | 29.35 | |
| GRPOBackbone=Qwen3-4B2026.06 | 29.14 | |
| VanillaBackbone=Qwen3-4B2026.06 | 28.31 | |
| ToolRLBackbone=Qwen3-4B2026.06 | 28.25 | |
| SFTBackbone=Qwen3-4B2026.06 | 25.12 | |
| PACTBackbone=Qwen3-1.7B2026.06 | 24.9 | |
| MatchTIRBackbone=Qwen3-1.7B2026.06 | 21.7 | |
| SFT→MatchTIRBackbone=Qwen3-1.7B2026.06 | 20.78 | |
| FTRLBackbone=Qwen3-1.7B2026.06 | 20.45 | |
| VanillaBackbone=Qwen3-1.7B2026.06 | 19.06 | |
| CHORDBackbone=Qwen3-1.7B2026.06 | 19 | |
| ToolRLBackbone=Qwen3-1.7B2026.06 | 18.52 | |
| GRPOBackbone=Qwen3-1.7B2026.06 | 18.08 | |
| SFTBackbone=Qwen3-1.7B2026.06 | 17.88 |