How Post-Training Shapes Biological Reasoning Models

How Post Training Shapes Biological Reasoning Models, led by Harvard University researchers with Google DeepMind and Google Research collaborators, studies more than 100 controlled multimodal biological reasoning checkpoints across DNA, RNA, and proteins using Qwen3-1.7B, Qwen3-4B, and Gemma 4 E2B backbones coupled to Evo2-1B, TranscriptFormer, or ESM-3 encoders. The paper shows that post-training stages are not additive: continued pre-training on a 200K/5K biology split from FineFineWeb improves downstream adaptation to biological language, supervised fine-tuning sharply raises in-domain accuracy but over-specializes and suppresses out-of-domain robustness, and reinforcement learning from a strong supervised checkpoint recovers transfer, especially OOD. In DNA pathway prediction, Qwen3-1.7B SFT lifts ID accuracy from 0.562 to 0.907 by 16 epochs while OOD peaks at 0.736 around 4 epochs and then falls to 0.687; in RNA target identification, Qwen3-4B reaches 0.911 ID accuracy but only 0.627 OOD before declining. RL reverses this pattern: Qwen3-4B RNA rises from 0.775/0.582 ID/OOD to 0.914/0.745 by 16 epochs, and protein function prediction with propagated GO-F1 reaches 0.956 OOD after RL. Under fixed post-training budgets, the best trade-off comes from brief SFT followed by larger RL, with mixed schedules such as 20% SFT and 80% RL reaching 0.9470 ID F1 and 0.9685 OOD F1 on proteins. The central takeaway is that biological reasoning improves most when compute, data, and adaptation capacity are allocated asymmetrically across stages rather than scaled uniformly.