A tool to catch reward hacking during RL training
rewardspy is a small tool for checking whether a rising reward score means a model is really improving or just exploiting the . It wraps an existing and watches training for early signs of . It tracks recent reward patterns, sudden drops in reward variation, imbalance between reward parts, changes in response length, shifts in reward slope, and GRPO group collapse.
The problem came up during GRPO training, where a higher reward made it hard to tell whether the policy was becoming better in a useful way. The library is still early and is being shared for technical .
Key points
- rewardspy wraps an existing and monitors training behavior.
- It is meant to separate real improvement from .
- It watches reward patterns, response length changes, reward slope shifts, and GRPO group collapse.
- The tool came from GRPO training experiments where reward increases were hard to trust.
- It is an early library and needs before serious use.