Cast a Wider Net: Coordinated Pass@K Policy Optimization for Code Reasoning

Researchers propose Coordinated Pass@K Policy Optimization (CPPO), a novel training method that improves code generation by having AI models explore multiple distinct algorithmic strategies simultaneously rather than sampling redundant solutions. Testing across competitive programming benchmarks shows significant performance gains, with improvements up to 27% on certain model configurations.

CPPO addresses a fundamental inefficiency in current code generation systems. When AI models generate multiple solutions through repeated sampling, they typically produce near-identical reasoning paths, wasting computational budget on redundant attempts. This problem becomes acute in competitive programming where diverse algorithmic approaches often exist for the same problem. The proposed method reframes pass@K generation as a joint exploration task: a planner module proposes K=4 distinct high-level strategies, and a solver attempts to implement each one. The training mechanism uses a multiplicative reward structure that credits only valid strategy tuples leading to verified success, encouraging genuine diversity. Empirical results across APPS, CodeContests, and LiveCodeBench-v6 demonstrate consistent improvements over baseline approaches including direct sampling, planning-only methods, and existing pass@K-optimized reinforcement learning. The largest measured gain reaches 27% on Qwen3.5-9B for LiveCodeBench-v6, achieving statistical significance. This advancement matters for AI development because it reveals inefficiencies in how computational resources are allocated during inference. As models scale and test-time computation becomes increasingly important, methods that eliminate waste through strategic diversity could substantially improve performance without requiring larger models or more samples. The research suggests that coordinating exploration across different solution approaches outperforms independent sampling, a principle with potential applications beyond code generation to other domains requiring diverse reasoning.

• →CPPO improves pass@4 performance by coordinating exploration across multiple distinct algorithmic strategies rather than sampling redundantly
• →The method uses a multiplicative planner reward that assigns credit only to valid strategy tuples achieving verified success
• →Testing shows statistically significant gains across multiple benchmarks, with maximum improvement of +0.16 (27% relative) on certain configurations
• →The approach addresses fundamental inefficiency where repeated sampling produces near-duplicate solutions wasting computational budget
• →Results suggest coordinated strategy exploration outperforms independent sampling under equivalent computational budgets