Rohan Paul (@rohanpaul_ai)T2·78 pts
- MoE design with 124B total and 5.5B active parameters spans medical image understanding and report generation
- RL rewards clinical meaning, abnormality localization, false positives and missed findings instead of report style imitation
- Ant claims leads over GPT-6.1 Sol, Claude Opus 5.5 and Gemini 3.8 Flash on AFUMED-Derm, AFUMED-Report, MIMIC-CXR and medical VQA