Medical Question Answering on HealthBench held-out heart-related n=500 seed 42
57AccuracyKimi-K2
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Kimi-K22026.04 | 57 | 52.6 | 72.6 | |
| GPT-OSS-120B2026.04 | 50.8 | 46.6 | 67.4 | |
| GRPO (COMPLEXITY)Backbone=Qwen3-14B2026.04 | 50.2 | 46 | 66.8 | |
| GRPO (HYBRID)Backbone=Qwen3-14B2026.04 | 49.8 | 45.4 | 66.5 | |
| MedGemma-27B2026.04 | 44.8 | 40.6 | 61.9 | |
| Gemma3-12B2026.04 | 44.2 | 39.8 | 61.3 | |
| Phi4-14B2026.04 | 44.2 | 39.8 | 61.3 | |
| Llama-4-Scout-17B2026.04 | 43.2 | 38.8 | 60.3 | |
| GRPO (RaR-IMPLICIT)Backbone=Qwen3-14B2026.04 | 41.2 | 37 | 58.4 | |
| Llama-3.3-70B2026.04 | 39.8 | 35.6 | 56.9 | |
| GRPO (RaR-EXPLICIT)Backbone=Qwen3-14B2026.04 | 39.6 | 35.4 | 56.7 | |
| MedGemma-4B2026.04 | 39.6 | 35.4 | 56.7 | |
| Llama-4-Maverick-17B2026.04 | 39 | 34.8 | 56.1 | |
| BaseBackbone=Qwen3-14B2026.04 | 36.2 | 32 | 53.2 | |
| MedGemma-1.5-4B2026.04 | 32.2 | 28.2 | 48.7 |