This repository provides the "Chain-of-Agents" (CoA) paradigm and the Agent Foundation Model (AFM), designed for end-to-end complex problem-solving within a single model by simulating multi-agent collaboration. It's targeted at researchers and developers looking to enhance LLM reasoning capabilities without complex external frameworks. The primary benefit is achieving state-of-the-art performance on benchmarks like GAIA through a novel distillation and reinforcement learning approach.

How It Works

The project utilizes a "Chain-of-Agents" distillation framework combined with agentic reinforcement learning. This approach trains a single model to dynamically activate tool agents and role-playing agents, mimicking multi-agent collaboration. This method is advantageous as it internalizes complex reasoning processes, leading to more efficient and capable end-to-end problem-solving compared to traditional multi-agent systems requiring manual orchestration.

Quick Start & Requirements

• Installation: Requires setting up a Conda environment (llama_factory or afm), installing dependencies via pip, and potentially cloning sub-repositories like apex and nsjail.
• Prerequisites: Python 3.10+, CUDA, deepspeed, swanlab, symeval, latex2sympy2, antlr4-python3-runtime, grpcio-status, protobuf, selenium. Building nsjail is necessary for the code execution sandbox.
• Data: Datasets for Web, MHQA, and Code agents need to be downloaded separately.
• Setup Time: Varies significantly based on dependency installation and dataset downloads, likely several hours.
• Links:
  • LLaMA-Factory (adapted)
  • verl (adapted)
  • nsjail

Highlighted Details

• Achieves SOTA performance: 55.3% Pass@1 on GAIA benchmark with a 32B AFM model.
• Demonstrates significant performance gains through test-time scaling (e.g., AFM-Pass@3 reaches 69.9% on GAIA).
• Fully open-sources data, models, and training/inference code for reproducibility.
• Supports dynamic activation of tool agents (Web, Code) and role-playing agents.

Maintenance & Community

The project is developed by OPPO PersonalAI Lab. It acknowledges contributions from the LLaMA-Factory and verl open-source projects. Further community engagement details (like Discord/Slack) are not explicitly mentioned in the README.

Licensing & Compatibility

The repository itself does not explicitly state a license. However, it heavily relies on and adapts code from LLaMA-Factory (Apache 2.0) and verl (MIT License). Users should verify the licensing of the specific models and datasets provided, as well as any combined usage implications.

Limitations & Caveats

The setup process involves numerous steps and dependencies, including building external tools like nsjail, which can be complex. Specific model checkpoints (e.g., AFM-WebAgent-32B-RL) may need to be downloaded separately for evaluation. The project is presented with a 2025 arXiv date, suggesting it may represent recent research rather than a mature, production-ready library.