KMnO4-zx/agentic-rl-lab

What it solves

This project provides a simplified, accessible way to reproduce and dismantle own-source LLM reinforcement learning (RL) algorithms. It aims to lower the GPU barrier and reduce code complexity compared to highly coupled engineering frameworks, making it easier for researchers and developers to study Agentic-RL.

How it works

Using the PyTRIO infrastructure, the repository implements a series of experimental records and tutorials. Each implementation includes a clear explanation of the algorithm's origin, the problem it solves, and its core variables, accompanied by runnable code for data, rewards, loss functions, and training loops, with experiment tracking via SwanLab.

Who it’s for

Algorithm engineers and researchers interested in studying and implementing frontier LLM RL and Agentic-RL methods without the overhead of complex infrastructure.

Highlights

• Broad Algorithm Coverage: Reproduces a wide range of methods including GRPO, OPD, OPSD, GSPO, DAPO, Search-R1, ReTool, ALFWorld, Vision GRPO, TEMPO, and AgentOPSD.
• Lighter Weight: Focuses on on-the-fly reproduction using simpler code and lower hardware requirements.
• Diverse Task Support: Includes implementations for mathematical reasoning (GSM8K), medical capabilities, multi-turn search, code-interleaved agents, and vision-based geometry problems (GeoQA).
• Environment Integration: Integrates with TextWorld for agent training in simulated household environments.