Worked on the inclusionAI/AReaL repository to enhance reinforcement learning stability by implementing the M2PO algorithm, focusing on reducing variance in off-policy training. The approach involved developing a dedicated loss function to constrain the second moment of importance weights and updating the loss mask logic to support these changes. Collaboration with Gemini guidance informed improvements to robustness and reliability for deployment. The work demonstrated proficiency in Python, algorithm development, and reinforcement learning, with all changes managed through Git-based workflows. This contribution addressed the challenge of stabilizing policy updates, resulting in more dependable training dynamics for reinforcement learning applications in production environments.