A director for LLM ensembles prone to groupthink

What it does

MASCOT builds multi-agent conversational systems—think emotional support circles or workplace meeting assistants—where each LLM speaker maintains a specific persona and a meta-level Director decides who speaks next and how. The framework trains both levels via reinforcement learning (GRPO with LoRA adapters) against small reward models, using rubric-based LLM judges to generate preference data. It is packaged as a full pipeline from synthetic episode generation through reward-model training to evaluation against an OpenAI-compatible vLLM server.

The interesting bit

The project treats social sycophancy—agents politely agreeing into uselessness—as a first-class optimization problem. A Director agent issues natural-language stage directions specifying speaker, action, and tone, then gets optimized with group-level trajectory rewards to keep the conversation complementary rather than redundant.

Key highlights

• Bi-level optimization: first align individual speakers to personas, then train a Director to orchestrate group dynamics.
• Persona fidelity is enforced via a Qwen3-0.6B reward model and GRPO; the README reports +14.1 consistency and +10.6 social contribution over baselines.
• The entire training stack uses LoRA adapters (0.187% trainable parameters, r=16) against frozen base weights referenced by Hugging Face name.
• Includes pre-built agent rosters for emotional support and workplace meetings, plus optional Big Five (OCEAN) personality profiles for simulated users.
• Ships with a self-contained pipeline: episode generation, rubric judging, preference-pair construction, reward modeling, GRPO/DPO training, and evaluation.

Caveats

• The documented pipeline centers on the Qwen3 model family and an OpenAI-compatible vLLM server; portability to other stacks is unexplored.
• Training defaults assume multi-GPU serving and leave hyperparameters buried in src/const.py.

Verdict

Worth a look if you are building multi-agent roleplay, therapy, or meeting-simulation systems and need a reproducible training recipe rather than ad-hoc prompt engineering. Skip it if you just want a simple multi-agent chat wrapper—this is a full RL training pipeline with heavy infrastructure assumptions.