Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties

Overview & Synopsis

Reinforcement learning for language models has moved from preference reward models to verifiable rewards, which sustain much longer optimization. Preference models are prone to reward hacking: the policy raises their score without improving response quality, or even while degrading it. Yet for open-ended tasks, where quality can neither be verified nor fully written down as rubrics, they remain valuable. We introduce Dynalize, which mitigates reward hacking online, during policy training. Periodically and asynchronously, a language model analyzes training rollouts and writes penalties for undesirable behaviors that the reward model appears to reward; a judge applies them to every rollout. In clinical counseling and creative writing, training on the reward model alone raises its score continuously but ultimately leads to a decline in external benchmarks; Dynalize sustains longer training and results in improved downstream performance. We find that our method suppresses patterns incentivized by preference models that, without penalties, spread to most responses, and even to the model's reasoning, which the reward model does not explicitly score. Furthermore, we show that its gains hold in multi-turn training, and even after we retrain the reward model on the discovered behaviors.