MAC: Multi-Agent Constitution Learning
A multi-agent system that writes its own rules outperforms human-crafted prompts without model fine-tuning.
A team of researchers including Rushil Thareja has introduced MAC (Multi-Agent Constitution Learning), a novel framework that automates the creation of rule-based "constitutions" to govern large language model (LLM) behavior. Traditional Constitutional AI relies on human experts to write natural language rules, but MAC deploys a network of specialized agents that accept, edit, or reject rule updates through structured collaboration. This addresses key limitations of existing LLM-based prompt optimizers, which require many labeled examples and suffer from diminishing returns as prompts grow. The team also developed MAC+, which improves performance by training agents on successful update trajectories.
In evaluations on a Personally Identifiable Information (PII) tagging task—a classification job where interpretability is critical and labeled data is limited—MAC outperformed recent prompt optimization methods by over 50%. The framework also demonstrated generalization to other agentic tasks like tool calling. Crucially, MAC achieved performance comparable to supervised fine-tuning and methods like GRPO (Group Relative Policy Optimization) without requiring any updates to the underlying model's parameters. The resulting constitutions are structured, human-readable, and auditable sets of rules, offering a transparent alternative to black-box prompt engineering or costly model retraining.
- Outperforms recent prompt optimization methods by over 50% on tasks like PII tagging.
- Generates human-readable, auditable rule sets without requiring model parameter updates (fine-tuning).
- Achieves performance comparable to supervised fine-tuning and GRPO, demonstrating a scalable alternative.
Why It Matters
Enables precise, interpretable control of LLMs for sensitive tasks without the cost and opacity of fine-tuning or manual prompt engineering.