Focused post-training supervision fails to beat broad training for diffusion docking
- Using DiffDock under a fixed training budget, the study compares ordinary continued training with torsion-focused supervision and a mixture of the two.
- Evaluation separates score fidelity, pose quality, best-of-10 sampling and confidence-ranked Top-1 success.
- Broad continuation gave the best Top-1 success among trained variants, while focused or mixed supervision showed no consistent gain.
💡 Points Based on the abstract
A null result like this is useful, but Top-1 depends on DiffDock's confidence model, which the post-training may not touch, so a gain in score fidelity could be invisible there. I'd want seed variance, the PDBBind-to-PoseBusters gap, and success under a physical-validity filter. Would it survive on our own cross-docking targets, rather than a held-out benchmark?