Democratic ICAI: Debating Our Way to Steering Principles from Preferences

Forum topic · 小凯 · 2026-06-30

Summary

Democratic ICAI is a new approach to preference-based AI alignment introduced by Kevin Kingslin, Anish Natekar, and Ashutosh Ranjan (arXiv:2606.28294). Standard Inverse Constitutional AI (ICAI) summarizes human preferences into natural-language principles, but its single-pass explanations miss the nuance behind complex decisions where multiple criteria interact. Democratic ICAI instead gathers multiple competing rationales through structured persona debate, producing a broader and more expressive account of the factors influencing each pairwise comparison. From these richer signals, the method derives clearer and more comprehensive steering principles, which are then used to guide decision modeling with LLM-based and decision-tree judges. Experiments on the creative preference benchmarks MuCE-Pref and LiTBench across multiple creative task categories show that Democratic ICAI yields more faithful preference structures. Compared with deliberative prompting and principle-based baselines, it improves average preference prediction across tasks and produces constitutions preferred by LLM annotators.

论文概要

研究领域: ML 作者: Kevin Kingslin, Anish Natekar, Ashutosh Ranjan 发布时间: 2026-06-26 arXiv: 2606.28294

Abstract

Preference-based alignment often struggles to capture the reasoning that underlies human judgments. Many evaluations rely on multiple interacting criteria, yet pairwise labels reveal only the final choice rather than the considerations that shape preferences. Inverse Constitutional AI (ICAI) improves interpretability in decision making by summarizing preferences into natural-language principles, but its single-pass explanations miss much of the nuance involved in complex decisions. We introduce Democratic ICAI, a novel approach that gathers multiple competing rationales through structured persona debate, offering a broader and more expressive account of the factors influencing each comparison. From these richer signals, we derive clearer and more comprehensive steering principles and use them to guide decision modeling via LLM-based and decision-tree judges.

Key Results

On the creative preference benchmarks MuCE-Pref and LiTBench, across multiple creative task categories, Democratic ICAI produces more faithful preference structures.
Relative to deliberative prompting and principle-based baselines, it improves average preference prediction across tasks.
The derived constitutions are preferred by LLM annotators.