1. Personas (expert / personality / …)

2. Response generators (critical / reasoning styles / …)

3. Discussion paradigms (memory / relay / …)

4. Decision protocols (voting / consensus / …)

We got 144+ configurations out of the box, all defined in one YAML file. Need more? Sub-class and drop your own component in.

You can also use any text-based HuggingFace dataset (we’ve tested on MMLU-Pro, WinoGrande, …); we have built-in evaluation pipelines so you can A/B test strategies without writing extra scripts, and it is easily customizable (pure Python) and Apache 2.0-licensed.

Repo: https://github.com/Multi-Agent-LLMs/mallm

Demo: https://mallm.gipplab.org/