Multi-Turn Tool Calling on BFCL v4 (val)
85Overall AccuracyRCTP-FT + RC-GRPO
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| RCTP-FT + RC-GRPOBackbone=Qwen2.5-7B2026.02 | 85 | 77.78 | 88.23 | 81.81 | 91.3 | |
| RCTP-FT + GRPOBackbone=Qwen2.5-7B2026.02 | 73.75 | 66.67 | 64.71 | 77.27 | 82.61 | |
| Opus-4.5Version=claude-opus-4-5-202511012026.02 | 61.25 | 55.56 | 65.22 | 64.71 | 59.09 | |
| GLM-4.7Version=glm-4.72026.02 | 58.75 | 55.56 | 58.82 | 54.55 | 65.22 | |
| Sonnet-4.5Version=claude-sonnet-4-5-202509292026.02 | 57.5 | 44.44 | 58.82 | 59.09 | 65.22 | |
| Gemini-3-ProVersion=gemini-3-pro-preview2026.02 | 53.75 | 55.56 | 64.71 | 36.36 | 60.87 | |
| GPT-5.2Version=gpt-5.2-2025-12-112026.02 | 50 | 38.89 | 47.06 | 45.45 | 65.22 | |
| SFT + GRPOBackbone=Qwen2.5-7B2026.02 | 48.75 | 55.56 | 35.29 | 50 | 52.17 | |
| RCTP-FT + RC-GRPOBackbone=LLaMA-3.1-8B2026.02 | 48.75 | 38.89 | 35.29 | 60.87 | 54.54 | |
| SFT + RC-GRPOBackbone=Qwen2.5-7B2026.02 | 46.25 | 50 | 35.29 | 50 | 47.82 | |
| RCTP-FT + GRPOBackbone=LLaMA-3.1-8B2026.02 | 46.25 | 33.33 | 35.29 | 54.55 | 56.52 | |
| SFT + GRPOBackbone=LLaMA-3.1-8B2026.02 | 35 | 38.89 | 17.67 | 40.91 | 39.13 | |
| SFT + RC-GRPOBackbone=LLaMA-3.1-8B2026.02 | 35 | 44.44 | 17.64 | 36.36 | 39.13 | |
| Base ModelBackbone=Qwen2.5-7B2026.02 | 11.25 | 11.11 | 17.65 | 4.54 | 13.04 | |
| Base ModelBackbone=LLaMA-3.1-8B2026.02 | 0 | 0 | 0 | 0 | 0 |