ThermoDPO Mitigates Reward Hacking via Manifold Drift Control
August 21, 2026
Manifold drift occurs when preference optimization in flow matching moves terminal samples off the pretrained data manifold. The proposed ThermoDPO objective uses temperature-controlled anchoring to minimize terminal displacement and prevent reward hacking.
HOW THIS AFFECTS YOU
●
researcherYou can apply this objective to stabilize flow-based generative models during preference alignment.