Rationale-Guided Policy Optimization Mitigates Reward Sparsity in Reasoning

Rationale-Guided Policy Optimization (RGPO) uses adaptive rationale scaffolding to provide training signals for difficult reasoning tasks. This framework prevents optimization stagnation by leveraging ground-truth rationale information to guide the agent through sparse reward environments.