Mechanistically guided fine-tuning, for PSF - MedGemma-4B (targeted + full) and LLaVA-RAD at n=12K multi-task