Worked on the volcengine/verl repository to address a critical issue affecting exploration and determinism in VLLM replicas. Focused on backend development and distributed systems using Python, the work involved fixing a bug where all replicas previously shared the same random seed, which limited exploration in deterministic mode. By differentiating seeds based on replica rank, the update enabled diverse random number generation trajectories, improving both stability and throughput in production workloads. The solution included comprehensive unit and end-to-end tests, expanded CI coverage, and updated documentation, ensuring robust quality assurance and alignment with project contribution guidelines for future development.