Conditional Equivalence of DPO and RLHF: Implicit Assumption, Failure Modes, and Provable Alignment

arXiv:2605.20834 · 2026-05-21 공개 · arXiv · PDF

policy-optimization rlhf preference-learning alignment dpo failure-modes constrained-preference-optimization margin-ranking

Abstract

Direct Preference Optimization (DPO) has emerged as a popular alternative to Reinforcement Learning from Human Feedback (RLHF), offering theoretical equivalence with simpler implementation. We prove this equivalence is conditional rather than universal, depending on an implicit assumption frequently violated in practice: the RLHF-optimal policy must prefer human-preferred responses. When this assumption fails, DPO optimizes relative advantage over the reference policy rather than absolute alignment with human preferences, leading to pathological convergence where policies decrease DPO loss while preferring dispreferred responses. We characterize when this assumption is violated, show the existence of an undesirable solution space, and prove that DPO and RLHF optimize fundamentally different objectives in such cases. To address this, we introduce Constrained Preference Optimization (CPO), augmenting RLHF with constraints for provable alignment. We further provide a geometric interpretation through soft margin ranking, revealing that DPO implements margin ranking with potentially negative targets. Our theoretical analysis establishes when DPOs' guarantees hold and provides solutions preserving simplicity with provable alignment. Comprehensive experiments on standard benchmarks demonstrate that CPO achieves state-of-the-art performance. Code is available at: https://github.com/visitworld123/CPO.

한국어 요약

📋 한 줄 요약

**[DPO 이론 / Alignment]** DPO–RLHF의 동치성이 "RLHF-optimal 정책이 인간 선호 응답을 선호한다"는 implicit 가정 하에서만 성립함을 증명·실패 모드 특성화, Constrained Preference Optimization(CPO)로 provable alignment 보장·SOTA 달성.

🎯 핵심 기여도

💡 핵심 아이디어

DPO와 RLHF의 동치성은 "RLHF-optimal 정책이 인간 선호 응답을 더 선호한다"는 implicit 가정에 묶여 있고, 가정이 깨지면 DPO는 relative advantage를 최적화하면서 dispreferred 응답을 선호하는 pathology에 빠지므로, RLHF에 제약을 더한 CPO로 absolute alignment를 provable하게 회복해야 한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: DPO의 이론적 한계를 정량적으로 노출하고 그 실패 모드를 형식화, RLHF 제약 보강(CPO)으로 provable alignment 회복, soft margin ranking 해석으로 DPO를 기하학적 이해. **한계**: 이론 가정의 검증 가능성·실제 데이터에서의 빈도 측정 필요, CPO 제약 설계가 도메인 의존, RLHF compute 오버헤드 유지.

🚀 실용적 활용