Medical Multimodal Reasoning on MMMU-Med, MMMU-Pro-Med-10, PMC-VQA, and OmniMedVQA (test)
58.96MMMU-Med AccuracyBase + SFT
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Base + SFTContinued Pre-Training (CPT) data source=None, CPT token count=–, Supervised Fine-Tuning (SFT) stage=LLaVA-OneVision 50K, Rule-based option extraction=True2026.05 | 58.96 | 41.61 | 61.4 | 85.37 | 61.84 | |
| PMC-InterCPTContinued Pre-Training (CPT) data source=PMC-InterCPT, CPT token count=9.63B, Supervised Fine-Tuning (SFT) stage=LLaVA-OneVision 50K, Rule-based option extraction=True2026.05 | 58.73 | 40.56 | 62.75 | 86.51 | 62.14 | |
| BIOMEDICAContinued Pre-Training (CPT) data source=Raw BIOMEDICA, CPT token count=13.65B, Supervised Fine-Tuning (SFT) stage=LLaVA-OneVision 50K, Rule-based option extraction=True2026.05 | 58.62 | 37.76 | 62.9 | 84.97 | 61.06 |