Welcome to Agent-R1! This repository is designed not just as a tool, but as a guide to understanding how to train Large Language Model (LLM) agents using Reinforcement Learning (RL).
If you've ever wondered how to move beyond simple prompting and supervised fine-tuning to agents that can learn from their interactions with tools and environments, you're in the right place.
This repository has the model checkpoint for grpo variant of qwen-2.5-1.5B model trained following the paper!
What are we trying to solve?
Training agents is hard. Unlike standard chatbots that just generate text, agents need to:
Reason about a problem.
Act by calling tools (search, calculators, code execution).
Observe the results of those actions.
Iterate until the task is solved.
Standard RLHF (Reinforcement Learning from Human Feedback) pipelines are typically designed for single-turn interactions:
Prompt -> Response -> Reward
But agents operate in a multi-turn loop. We need a framework that treats the entire interaction trajectory as a training episode.
Built on VERL
Agent-R1 is built on top of VERL (VolcEngine Reinforcement Learning). VERL provides the robust, scalable infrastructure needed for distributed RL training (handling the "heavy lifting" of PPO/GRPO across GPUs).
Agent-R1 adds the "Agent" logic on top of this infrastructure. Think of VERL as the engine, and Agent-R1 as the car designed for a specific type of race (agentic tasks).
Key Concepts & Contributions
To make RL work for agents, Agent-R1 introduces two critical concepts that you'll see annotated throughout the codebase:
1. Algorithm: Agent vs. LLM RL
Reinforcement learning for Agents differs significantly from LLM (Chatbots, Reasoners). The key distinction lies in the fact that: a complete Agent trajectory typically involves multi-turn interaction, requiring multiple tool calls to solve user queries. Below, we formalize this distinction using the Markov Decision Process (MDP) framework.
In the context of LLM:
State: Simply the sequence of the input prompt and all generated text so far.
Action: Selecting the next token from the vocabulary to add to the sequence.
Transitions: Straightforward addition of the selected token to the existing sequence.
Rewards: Typically only provided at the end of the sequence generation.
For Agents, the components are more complex:
State: Includes not only the input and generated text, but also all tool responses from previous interactions.
Action: Still involves selecting the next token, but some tokens can trigger tool calls.
Transitions:
Regular tokens: Simply add to the sequence like in traditional LLM training.
Tool-triggering tokens: Cause external tool execution that produces responses, introducing significant stochasticity into state transitions unlike the deterministic nature of standard LLM generation.
Rewards: Can be provided at multiple points:
After each tool call, which naturally creates process-level rewards based on the quality and effectiveness of tool usage.
At the end of the complete interaction based on overall task completion.
Formal Definition:
To better understand the difference, we can define the probability of a trajectory:
$C_j$ is the result of the $j$-th tool call and $m$ is the number of tool calls
$a_t$ is the token selected from the vocabulary
$t_j$ is the number of token responses between the $j-1$th and $j$-th tool calls, $0<t_1+t_2+...+t_m<t$
This richer reinforcement learning framework allows Agent-R1 to train LLMs that learn effective strategies for when and how to use tools across multi-turn interactions. By optimizing over entire trajectories rather than single responses, we can apply algorithms like PPO and GRPO to develop agents that reason effectively before taking actions.
2. Advantage & Loss Masks
This is the technical "secret sauce". In a multi-turn trajectory, the context window contains a mix of:
User instructions (Input)
Model generations (Thoughts/Actions)
Tool outputs (Environment Observations)
We only want to update the model based on its own actions. We don't want to calculate gradients for the tool outputs (the model didn't generate them, the environment did).
Agent-R1 implements sophisticated masking:
Loss Mask: Ensures we only calculate loss on the tokens the model generated.
Advantage Mask: Ensures that the "credit" for a good outcome is properly attributed to the specific actions the model took, skipping over the deterministic tool outputs.
Simplicity First
Inspired by projects like Nanochat, we aim for code readability and ease of understanding.
Focused Algorithms: We strictly support PPO (Proximal Policy Optimization) and GRPO (Group Relative Policy Optimization). We've removed support for older or less stable algorithms (like REINFORCE, ReMax) to keep the codebase clean.
Tutorial Style: The code is heavily commented to explain why we are doing things, not just what is happening.
Running the model
We provide the main scripts to help you get started, depending on your environment and needs.
3. Inference
Once training is complete, you can interact with your trained agent using the inference script. For more detailed information, please refer to docs/inference/inference.md.
bash
1# Run inference (interactive chat)2bash inference.sh
This script will:
Convert your training checkpoints to Hugging Face format.
Start a vLLM server to host the model.
Launch an interactive chat interface where you can test the agent.
You can also customize the paths if your checkpoints are in a different location:
