Training a general LLM-based reasoning model for agent-compiled knowledge refinement that evolves any KB with multi-turn interaction history via RL. - View it on GitHub