langfengQ/verl-agent

# Technical Summary Implements a step-independent multi-turn rollout mechanism that decouples per-step input structures from interaction history, enabling customizable memory management and keeping context length constant across long-horizon tasks. Provides diverse RL algorithms (GiGPO, GRPO, PPO, DAPO, RLOO) with support for both text and vision-language models across multiple agent environments (ALFWorld, WebShop, Sokoban, AppWorld). Built on the veRL framework and includes modular memory managers, parallelized gym environments, and group-based RL for scalable agent training up to 50+ step trajectories.

1,668 stars.