Shen, S., Hou, Z., Sun, J., Zhang, Y., Lin, W., Yin, G., Huang, Q., & Wang, S. (2026). On-Policy Parameter Update Direction Underlies Generalization in LLM Post-Training. https://omanscience.com/en/articles/on-policy-parameter-update-direction-underlies-generalization-in-llm-post-training