The Hidden Link Between RLHF and Contrastive Learning

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Lv, Xufei, Chen, Kehai, Sun, Haoyuan, Bai, Xuefeng, Zhang, Min, Liu, Houde
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914090057728000
author Lv, Xufei
Chen, Kehai
Sun, Haoyuan
Bai, Xuefeng
Zhang, Min
Liu, Houde
Chen, Kehai
author_facet Lv, Xufei
Chen, Kehai
Sun, Haoyuan
Bai, Xuefeng
Zhang, Min
Liu, Houde
Chen, Kehai
contents Alignment of large language models (LLMs) with human values has recently garnered significant attention, with prominent examples including the canonical yet costly Reinforcement Learning from Human Feedback (RLHF) and the simple Direct Preference Optimization (DPO). In this work, we demonstrate that both RLHF and DPO can be interpreted from the perspective of mutual information (MI) maximization, uncovering a profound connection to contrastive learning. Within this framework, both RLHF and DPO can be interpreted as methods that performing contrastive learning based on the positive and negative samples derived from base model, leveraging the Donsker-Varadhan (DV) lower bound on MI (equivalently, the MINE estimator). Such paradigm further illuminates why RLHF may not intrinsically incentivize reasoning capacities in LLMs beyond what is already present in the base model. Building on the perspective, we replace the DV/MINE bound with the Jensen-Shannon (JS) MI estimator and propose the Mutual Information Optimization (MIO). Comprehensive theoretical analysis and extensive empirical evaluations demonstrate that MIO mitigates the late-stage decline in chosen-likelihood observed in DPO, achieving competitive or superior performance across various challenging reasoning and mathematical benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2506_22578
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The Hidden Link Between RLHF and Contrastive Learning
Lv, Xufei
Chen, Kehai
Sun, Haoyuan
Bai, Xuefeng
Zhang, Min
Liu, Houde
Chen, Kehai
Machine Learning
Artificial Intelligence
Alignment of large language models (LLMs) with human values has recently garnered significant attention, with prominent examples including the canonical yet costly Reinforcement Learning from Human Feedback (RLHF) and the simple Direct Preference Optimization (DPO). In this work, we demonstrate that both RLHF and DPO can be interpreted from the perspective of mutual information (MI) maximization, uncovering a profound connection to contrastive learning. Within this framework, both RLHF and DPO can be interpreted as methods that performing contrastive learning based on the positive and negative samples derived from base model, leveraging the Donsker-Varadhan (DV) lower bound on MI (equivalently, the MINE estimator). Such paradigm further illuminates why RLHF may not intrinsically incentivize reasoning capacities in LLMs beyond what is already present in the base model. Building on the perspective, we replace the DV/MINE bound with the Jensen-Shannon (JS) MI estimator and propose the Mutual Information Optimization (MIO). Comprehensive theoretical analysis and extensive empirical evaluations demonstrate that MIO mitigates the late-stage decline in chosen-likelihood observed in DPO, achieving competitive or superior performance across various challenging reasoning and mathematical benchmarks.
title The Hidden Link Between RLHF and Contrastive Learning
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2506.22578