CoRe-Code: Collaborative RL Framework for Code Generation

A new framework called CoRe-Code (Collaborative Reinforcement Code) addresses limitations in LLM-based code generation by introducing role-specialized agents with collaborative reinforcement learning. Current autoregressive methods lack global planning, often producing locally coherent but globally suboptimal code. Approaches like Chain-of-Thought and multi-agent systems improve planning but suffer from poor role specialization and coordination. CoRe-Code employs a Planner-Coder paradigm: the Planner generates high-level plans, and the Coder executes them. A collaboration-aware reinforcement learning mechanism enhances inter-agent coordination. The framework aims to produce more accurate and efficient code, particularly for complex tasks. The research is published as arXiv:2605.24812v1.