Tool-use alignment on FA-Bench 500 queries
91TCRToolRLA
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ToolRLALatency=1.6s2026.03 | 91 | 14 | 2.8 | 96 | 0.8 | |
| ReAct + GRPO (coarse)Latency=1.9s2026.03 | 82 | 21 | 3.1 | 74 | 6.5 | |
| ReAct + GRPO (additive)Latency=1.8s2026.03 | 80 | 22 | 3 | 75 | 6.2 | |
| ReAct + PPO (binary)Latency=2.1s2026.03 | 76 | 26 | 3.4 | 70 | 8 | |
| ReAct + SFT onlyLatency=2.4s2026.03 | 68 | 38 | 3.9 | 65 | 10.5 | |
| Multi-Model PipelineLatency=2.8s2026.03 | 62 | 38 | 4.2 | 61 | 12 |