OpenMLRL

Multi-LLM
Reinforcement Learning

We develop open-source reinforcement learning frameworks and tools for multi-LLM systems.

01 / CoMLRL

Cooperative Multi-LLM
Reinforcement Learning

CoMLRL is an open-source library for training multiple language model agents with reinforcement learning and preference learning.

  • Multi-Agent
  • Multi-Turn
  • Decentralized Agents
Documentation GitHub
Multiple agents, joint completionThree language model agents connect to a shared circular composition representing their joint completion.AGENT 01AGENT 02AGENT 03Joint Completion

About OpenMLRL

We are a group of researchers working on reinforcement learning for collaborative language model systems. Our work covers multi-agent reinforcement learning, preference learning, multi-turn interaction, and decentralized systems.

01

Cooperative learning

Training language model agents to coordinate their responses and solve tasks together.

02

Preference learning

Optimizing multi-LLM systems through pairwise preferences.

03

Multi-turn interaction

Improving multi-LLM system performance through multi-turn interaction with the environment.