MASCOT

Why we included this project

Teams building multi-agent LLM systems that must hold distinct roles rather than collapse into one generic assistant will find a lot to study here. MASCOT is a research framework, backed by an EMNLP paper, that tackles two failure modes common in multi-agent setups: agents drifting out of their assigned personas, and agents piling on redundant, agreeable dialogue. It does this with a two-stage pipeline, first aligning each speaker to a strict persona via GRPO against a trained reward model, then optimizing a director agent that picks who speaks next and issues natural-language directives for tone and action. The codebase is a self-contained pipeline covering episode generation, rubric judging, preference-data construction, reward modeling, and GRPO/DPO training against an OpenAI-compatible vLLM server, with LoRA-only training to keep resource use low. It is primarily a reference implementation for researchers and engineers reproducing or extending the method, not a turnkey product, so expect to adapt it to your own rosters and datasets rather than deploy it out of the box.