Abstract
Reward models in reinforcement learning suffer from oversensitivity issues where they assign different scores to equally good responses, leading to poor policy learning, but this can be mitigated through discretization techniques that maintain discriminative ability while reducing oversensitivity.
Despite their widespread use, the role of reward models in shaping reinforcement learning is poorly understood. Reward models offer a tempting promise: they automatically estimate response quality in the absence of verifiers or human judges. Unlike "verifiable rewards" which typically produce binary scores, reward models typically produce continuous scores, allowing them to be sensitive to fine-grained differences in responses. However, we show this apparent strength is a serious weakness: many popular reward models are oversensitive, assigning different scores to equally good responses. Theoretically, we show that seemingly perfect reward models can be highly oversensitive; empirically, this oversensitivity can lead to bad policies. In place of existing notions of "reward model accuracy," we propose evaluating reward models using distinct measures of "discriminative ability" and "specificity" (the complement of oversensitivity). As a solution, we describe a training-free algorithm that uses Monte Carlo dropout on any neural reward model to produce discrete reward clusters. Theoretically, we prove there exist discretizations that reduce oversensitivity at minimal expense of discriminative ability; empirically we show, in both controlled and natural RL settings, that discretizing rewards leads to less reward hacking and better policies than training on the original rewards.
Community
Reward models are an essential tool in language model alignment and post-training. But we don't really understand what makes a given reward model suitable as a teacher for reinforcement learning. Some reward models that are near-perfect at imitating human preferences are very bad at guiding RL.
We offer a potential explanation: reward model oversensitivity. Turns out that even the strongest-known reward models see large differences within groups of equally-good responses. This is a bad thing. We propose a way to evaluate reward models (by separately measuring "specificity" -- the opposite of oversensitivity -- and "discriminative ability") and a new algorithm that, on-the-fly, can make any existing reward model less oversensitive.
For every reward model we tried,, discretization either maintains or improves how well it can teach a student model via RL. This is evidence that specificity/oversensitivity might be an answer to the question of "what makes a reward model a bad teacher for RL".
Get this paper in your agent:
hf papers read 2606.21795 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper