v1.0.0

CoMLRL v1.0.0 provides the code used in LLM Collaboration with Multi-Agent Reinforcement Learning (AAAI 2026). The paper introduces Multi-Agent Group Relative Policy Optimization (MAGRPO) to train LLMs to collaborate on writing and coding tasks.

Note

Use v1.0.0 to reproduce the paper. For new projects, see the latest CoMLRL release and the current documentation.

Core Features

1. MAGRPO training. Configure the number of agents and interaction turns in MAGRPOTrainer. With one agent, MAGRPO reduces to GRPO. For multi-turn tasks, an external transition function supplies the context for the next turn.
2. Task-specific rewards. Define how to score the agents' responses and combine multiple reward functions with configurable weights.
3. Writing and coding examples. The tagged examples cover arXiv abstract expansion, TL;DR summarization, collaborative story writing, and LeetCode-style function and test generation.

Changelog

• Introduced MAGRPOTrainer with support for configurable agent counts and single- or multi-turn training.
• Released matching HumanEval and CoopHumanEval experiments with results for single-agent GRPO and single- and multi-turn MAGRPO.