Implicit Safety Alignment from Crowd Preferences

Qian Lin, Daniel S. Brown

arXiv:2605.21822 · 2026-05-23 공개 · arXiv · PDF

llm-training reward-modeling rlhf safety-alignment preference-learning safe-rl safety-costs crowd-preferences

Abstract

Reinforcement Learning from Human Feedback (RLHF) can reveal implicit objectives such as safety considerations that go beyond task completion. In this work, we focus on the common safety criteria embedded in crowd preference datasets, where different users may express distinct preferences or objectives, yet follow similar safety principles. Our aim is to discover shared safety criteria from crowd preferences and then transfer them to downstream RL tasks to regularize agent behavior and enforce safety. We first show that direct reward combination-optimizing a preference-learned reward model together with downstream task rewards-has inherent limitations. Motivated by this, we propose Safe Crowd Preference-based RL, a hierarchical framework that extracts safety-aligned skills from crowd preferences and composes them via a high-level policy to safely solve downstream tasks. Experiments across safe RL environments and a preliminary LLM-style task with diverse user goals and shared safety constraints demonstrate that our approach substantially lowers safety costs without access to explicit safety rewards, while achieving task performance comparable to oracle methods trained with ground-truth safety signals.

한국어 요약

한 줄 요약

**[RLHF / 안전 정렬]** Safe Crowd Preference-based RL이 군중 선호 데이터에서 공유 안전 기준을 추출해 hierarchical skill로 구성, 명시적 safety reward 없이도 oracle 수준 task 성능 유지하며 safety cost 대폭 감소.

핵심 기여도

핵심 아이디어

다양한 사용자 선호가 공유하는 implicit 안전 원칙을 reward 결합으로 단순 합치는 대신, hierarchical 구조로 안전 skill과 task 정책을 분리·결합하면 explicit safety reward 없이도 안전과 task 성능을 동시 달성할 수 있다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: RLHF에서 safety를 implicit objective로 추출하는 새 패러다임, hierarchical 분해로 safety·task 분리, oracle 수준 성능을 explicit signal 없이 달성. **한계**: 군중 선호의 안전 합의가 약할 때의 추출 한계, LLM-style task가 preliminary 수준, hierarchical 구조 hyperparameter 부담.

실용적 활용