SLIME: Stabilized Likelihood Implicit Margin Enforcement for Preference Optimization

TL;DR

SLIME introduces a novel alignment method for large language models that stabilizes preference learning, prevents unlearning, and improves generation quality by combining anchoring, stabilization, and dual-margin constraints.

Contribution

It proposes a new reference-free objective that decouples preference optimization from generation quality, addressing critical issues in existing methods.

Findings

SLIME outperforms state-of-the-art baselines in alignment tasks.

It maintains higher generation stability compared to previous methods.

The approach effectively prevents unlearning and formatting collapse.

Abstract

Direct preference optimization methods have emerged as a computationally efficient alternative to Reinforcement Learning from Human Feedback (RLHF) for aligning Large Language Models (LLMs). Latest approaches have streamlined the alignment process by deriving implicit reward functions, yet they often suffer from a critical objective mismatch: optimizing the relative margin between chosen and rejected responses does not guarantee the preservation of the chosen response's absolute likelihood. This can lead to unlearning, where the model degrades the probability of high-quality outputs to satisfy margin constraints, and formatting collapse caused by the over-penalization of rejected sequences. In this work, we introduce SLIME (Stabilized Likelihood Implicit Margin Enforcement), a reference-free alignment objective designed to decouple preference learning from generation quality. SLIME incorporates a three-pronged objective: (1) an anchoring term to maximize the likelihood of preferred responses; (2) a stabilizing penalty that prevents the probabilities of rejected tokens from collapsing to zero; and (3) a dual-margin mechanism that combines hard and soft constraints for precise boundary shaping. Our results demonstrate that SLIME achieves superior performance compared to state-of-the-art baselines while maintaining higher generation stability.