# MAC: Multi-Agent Constitution Learning

Source: [MAC: Multi-Agent Constitution Learning](<https://www.sota2.com/research/paper/69ba317b0636e5bf4afc072c>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-19T05:00:05.030Z

Authors: Rushil Thareja, Gautam Gupta, Francesco Pinto, Nils Lukas

Published: 2026-03-16T22:42:07.000Z

## Abstract

Constitutional AI is a method to oversee and control LLMs based on a set of rules written in natural language. These rules are typically written by human experts, but could in principle be learned automatically given sufficient training data for the desired behavior. Existing LLM-based prompt optimizers attempt this but are ineffective at learning constitutions since (i) they require many labeled examples and (ii) lack structure in the optimized prompts, leading to diminishing improvements as prompt size grows. To address these limitations, we propose Multi-Agent Constitutional Learning (MAC), which optimizes over structured prompts represented as sets of rules using a network of agents with specialized tasks to accept, edit, or reject rule updates. We also present MAC+, which improves performance by training agents on successful trajectories to reinforce updates leading to higher reward. We evaluate MAC on tagging Personally Identifiable Information (PII), a classification task with limited labels where interpretability is critical, and demonstrate that it generalizes to other agentic tasks such as tool calling. MAC outperforms recent prompt optimization methods by over 50%, produces human-readable and auditable rule sets, and achieves performance comparable to supervised fine-tuning and GRPO without requiring parameter updates.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Private Information Tagging on ECHR (test)](<https://www.sota2.com/research/sota/private-information-tagging-on-echr-test.md>): Private Information Tagging; ECHR (test); F1 Score: 67.3.

- [Private Information Tagging on PUPA subset (test)](<https://www.sota2.com/research/sota/private-information-tagging-on-pupa-subset-test.md>): Private Information Tagging; PUPA subset (test); F1 Score: 45.5.

- [Private Information Tagging on MACCROBAT (test)](<https://www.sota2.com/research/sota/private-information-tagging-on-maccrobat-test.md>): Private Information Tagging; MACCROBAT (test); F1 Score: 0.267.

- [Tool-calling on BFCL Generalization (test)](<https://www.sota2.com/research/sota/tool-calling-on-bfcl-generalization-test.md>): Tool Calling; BFCL Generalization (test); Final Accuracy: 87.5.

- [Private Information Tagging on Legal (test)](<https://www.sota2.com/research/sota/private-information-tagging-on-legal-test.md>): Private Information Tagging; Legal (test); Precision: 75.93.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.15968>)

[PDF](<https://arxiv.org/pdf/2603.15968.pdf>)
