This repository is an "awesome list" curating open-source projects that train Large Language Model (LLM) agents using Reinforcement Learning (RL). It targets researchers and developers in the LLM agent space, providing a structured overview of projects focusing on RL frameworks, algorithms, reward mechanisms, and environments.

How It Works

The project compiles a list of repositories based on code analysis, specifically identifying agents that feature multi-turn interactions or tool usage. It categorizes projects by their RL framework, RL algorithm, reward type (e.g., external verifier, model-based, rule-based), and the tasks they address (e.g., math, code, web browsing, QA). This curated data aims to help users understand the technical choices made in these projects.

Highlighted Details

• Extensive categorization of RL-based LLM agents across various domains including web interaction, GUI control, tool usage, text games, and code generation.
• Detailed breakdown of RL algorithms (PPO, GRPO, DPO, etc.) and reward types employed by different agent projects.
• Includes a comprehensive list of environments relevant to LLM agent training, such as Mind2Web, AgentBench, and TextWorld.
• Projects are tagged with their primary tasks, RL frameworks, and tool usage capabilities.

Maintenance & Community

The project welcomes contributions and encourages users to report errors or omissions via issues or pull requests. It highlights contributions from various academic institutions and industry labs, including Microsoft Research, Alibaba, Tsinghua University, and HuggingFace.

Licensing & Compatibility

The repository itself is a curated list and does not appear to have a specific license. However, the linked projects likely have their own licenses, which would need to be checked individually for compatibility with commercial or closed-source use.

Limitations & Caveats

The project's data is derived from code analysis using GitHub Copilot Agent, which may lead to unfaithful cases or omissions despite manual review. The "Date" column appears to be a custom metric or versioning, not a standard release date.