Skip to content- Sources: primary
- Summary: The authors formalize how traits evolve during RL post-training and recover the Price equation from evolutionary biology. They call the amplification or suppression of non-targeted traits trait hitchhiking and show the Price estimates closely track sampled trait dynamics in models trained with GRPO.
- Why it matters: It gives post-training teams a way to predict which side behaviors an RL run will amplify before they show up in evals.
- Background: The Price equation splits a change in a population's average trait into a part driven by covariance with fitness and a part from transmission. In RL with verifiable rewards, reward plays the role of fitness, so any trait correlated with rewarded completions can spread. GRPO is a policy gradient method that scores each completion against others sampled for the same prompt.