Skip to content

MoA

Layered LLM agents that pool multiple model outputs

Visit official site

What MoA is

The method is described as a Mixture of Agents approach that layers multiple LLM agents to draw on their collective strengths.

What you can do with it

  • Run a layered set of LLM agents
  • Hold multi-turn dialogue with consistent context
  • Adjust aggregator and reference model settings

Application scenarios

  • Reproduce the AlpacaEval 2.0 result with scripts
  • Compare FLASK dimensions such as harmlessness and robustness
  • Build a custom multi-turn dialogue setup

Pros & cons

✓ What we like

  • Multi-turn dialogue keeps context consistent
  • A 7.6 percent absolute gain is listed

! What to watch out for

  • A Together API key is set as an environment variable
  • The demo uses predefined reference models

FAQ

What score is reported?

65.1 percent on AlpacaEval 2.0 with open models only.

What do I configure?

Aggregator, reference model and temperature.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review