Tool-calling for Clinical Question Answering on AgentBench FHIR (val)
76.7ScoreQWEN3-8B + RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| QWEN3-8B + RLRL Recipe=GRPO + clip-higher + DAPO-filter, rollouts=52026.05 | 76.7 | 79.7 | |
| GLM-52026.05 | 58.7 | — | |
| GEMINI 3 FLASH2026.05 | 51.9 | — | |
| O4-MINI2026.05 | 46.9 | — | |
| QWEN3-32Brollouts=52026.05 | 39.3 | 68.6 | |
| QWEN3-14Brollouts=52026.05 | 32.5 | 53.8 | |
| QWEN3-8Brollouts=52026.05 | 19.7 | 47.2 | |
| QWEN3-4Brollouts=52026.05 | 11.1 | 29.5 |