Mixture-of-Agents (MoA) is a framework for enhancing LLM performance by orchestrating multiple specialized agents. It targets researchers and developers seeking to improve response quality and achieve state-of-the-art results, particularly with open-source models, by leveraging collective intelligence.

How It Works

MoA employs a layered architecture where each layer consists of multiple LLM agents. These agents process input, and their outputs are aggregated, potentially through multiple refinement layers, to produce a final, superior response. This approach allows MoA to harness the diverse strengths of different models, leading to significant performance gains over single, even larger, models.

Quick Start & Requirements

• Install: pip install together
• API Key: Requires TOGETHER_API_KEY environment variable.
• Demo: pip install -r requirements.txt then python bot.py for CLI demo.
• Evaluation: Requires TOGETHER_API_KEY and OPENAI_API_KEY. Setup involves installing dependencies for AlpacaEval, MT-Bench, and FLASK.
• Links: Overview, Quickstart, CLI Demo, Evaluation

Highlighted Details

• Achieves 65.1% on AlpacaEval 2.0 using only open-source models, outperforming GPT-4 Omni (57.5%).
• Outperforms Qwen1.5-110B-Chat and GPT-4 Omni on specific FLASK evaluation dimensions like correctness and factuality.
• Supports multi-layer refinement for improved response quality.
• Includes scripts to reproduce results for AlpacaEval, MT-Bench, and FLASK benchmarks.

Maintenance & Community

• Developed by Together AI.
• Acknowledges contributions from Meta AI, Mistral AI, Microsoft, Alibaba Cloud, and DataBricks for base models.
• Credits LMSYS and KAIST AI for evaluation benchmarks.

Licensing & Compatibility

• Licensed under Apache 2.0.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The project relies on the Together API for model inference, which may incur costs and requires an API key. While it showcases strong performance on benchmarks, real-world applicability may vary based on specific use cases and the chosen agent configurations.