Breaking $\textit{Winner-Takes-All}$: Cooperative Policy Optimization Improves Diverse LLM Reasoning

In three linesGCPO (Group Cooperative Policy Optimization) replaces competitive rollout optimization with team-level credit assignment. Rollouts are rewarded by contribution to valid solution coverage (determinant volume over semantic embeddings), not individual accuracy. Results: improved reasoning accuracy and solution diversity across benchmarks.

Summary generated by Claude — human-verified