ThinkPrior: Zero-Rollout Difficulty Priors for Cold-Start Prompt Selection in RLVR

A revised arXiv paper proposes ThinkPrior, a method for choosing cold-start prompts in reinforcement learning with verifiable rewards without running any rollouts first. The authors observe that the KL-free reward-advantage term used in group relative policy optimization depends on how much reward varies within a group of rollouts, which collapses when a prompt is uniformly easy or hard. They use difficulty priors to pick prompts that are likely to produce useful within-group variation, aiming to make early training more efficient.