Awesome AI Agents › Multi-Agent Frameworks

hello-diana/MASCOT

⭐ 595 Python repository created 2026-02-03

MASCOT is a research framework for multi-agent conversational companions that hold distinct personas while talking to a user together. It addresses two failure modes the authors observe in multi-agent LLM chat: persona collapse, where agents drift back into a generic assistant voice, and social sycophancy, where agents echo one another instead of adding complementary content. The answer is a two-stage bi-level optimisation. Stage one aligns each speaker agent to its persona: eight candidate replies are generated per context, an LLM judge scores them against rubrics, preference pairs are built from the scores, a small Qwen3-0.6B reward model is trained on those pairs, and the Qwen3-8B speaker policy is optimised with GRPO against that reward plus format and conciseness terms. Stage two optimises the group: a director agent chooses who speaks next and issues a natural-language directive naming speaker, action and tone, whole trajectories are scored by a group-level reward model, and the director is trained with GRPO over trajectory rewards; an offline DPO trainer over branching episodes remains as an alternative. Only LoRA adapters are trained, about 0.187 percent of parameters, with base weights referenced by Hugging Face name. The repository is a self-contained pipeline with separate entry points for episode generation, rubric judging, preference data construction, reward modelling, GRPO and DPO training and evaluation, all running against an OpenAI-compatible vLLM server. Reported gains are 14.1 on persona consistency and 10.6 on social contribution, with 69 percent human preference head to head, evaluated on Empathetic Dialogues and QMSum meetings plus out-of-domain ESConv. It is aimed at researchers working on emotional support dialogue, meeting assistants and reinforcement learning for multi-agent interaction quality.

https://github.com/hello-diana/MASCOT

multi-agentllm-agentsgrporlaifpersonadialogue-systemsresearch-paperemnlpnlp