StaRPO: Stability-Augmented Reinforcement Policy Optimization

StaRPO is a stability-augmented reinforcement learning framework for complex reasoning. It combines task rewards with two lightweight process-aware signals: the Autocorrelation Function for local step-to-step coherence and Path Efficiency for global goal-directedness. Experiments across four reasoning benchmarks show improvements in both final-answer accuracy and logical stability.

📄 In Proceedings of the Fortieth Annual Conference on Neural Information Processing Systems (NeurIPS 2026).
🔗 Paper Link (arXiv:2604.08905)