Skip to content
프로젝트

AQ-MedAI/MrlX has 224 GitHub stars and 13 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-22 18:17 UTC. Last push 2026-01-19.

이 수치를 측정하는 방법

측정된 스타 증가

in 24h
0
in 7 days
0
in 15 days
기록 수집 중
in 1 month
기록 수집 중

서로 겹치는 기간의 순증가량이며 일별 합계가 아닙니다. 측정값이 없으면 비워 둡니다. 스타는 관심도를 나타내며 품질을 보장하지 않습니다.

소개

MrlX: A Multi-Agent Reinforcement Learning Framework

학습

README.md

Original

MrlX

Multi-Agent Reinforcement Learning Framework

In MrlX, Agent A and Agent B operate as independent agents, communicating via a message queue that enables cross-agent API calls, abstracts internal logic into external requests, and supports multi-turn interactions, inference result sharing, and collaborative decision-making.

At runtime, Agent A initiates multi-turn dialogue generation, while Agent B also engages in multi-turn responses. The coordination module evaluates both agents’ dialogues, calculates bilateral rewards, and drives iteration through the message queue. Each agent maintains a complete train–infer loop: the Data Buffer manages training samples, the SGLang Router schedules inference tasks, and Megatron executes model training, forming a “Generate → Train → Sync” flywheel mechanism.

Training data flows from the Data Buffer into Megatron, where updated weights are synchronized back to the inference service. This enables efficient knowledge transfer and continuous co-evolution between agents, transcending single-task limitations and allowing multi-agent systems to improve decision-making capabilities in dynamic environments.

Table of Contents

Architecture-Overview

framework

Module Description

  • training (Megatron): Responsible for the main training process; reads data from the Data Buffer, trains the model, and synchronizes updated parameters to the Rollout module
  • rollout (SGLang + router): Generates new data (including reward calculation and verification) and stores it in the data buffer
  • data buffer: Serves as a bridge between training and inference, managing prompt initialization, custom data loading, and rollout-generated content
  • custom rollout generation: Implements custom data generation logic, tailoring multi-turn interaction strategies, output formats
  • message queue: Transfers multi-turn interaction information between Agent A and Agent B, supports cross-agent API communication, task distribution, and state synchronization, and drives the iterative loop

Use-Cases

MrlX-TakesTwo

See MrlX-TakesTwo

MrlX-DeepResearch

See MrlX-DeepResearch

Acknowledgements

  • Special thanks to the following projects & communities: slime, Miles, SGLang, Megatron-LM, and others.

관련 프로젝트 살펴보기

불러오는 중…