LLM Alignment as Retriever Optimization: An Information Retrieval Perspective

TL;DR

This paper introduces a novel approach to LLM alignment by applying principles from Information Retrieval, specifically using a retriever-reranker framework to improve alignment quality effectively and simply.

Contribution

The paper presents LarPO, a new IR-inspired method for LLM alignment that simplifies the process and demonstrates significant improvements over existing approaches.

Findings

LarPO achieves 38.9% improvement on AlpacaEval2.

LarPO achieves 13.7% improvement on MixEval-Hard.

The IR-based framework effectively enhances LLM alignment quality.

Abstract

Large Language Models (LLMs) have revolutionized artificial intelligence with capabilities in reasoning, coding, and communication, driving innovation across industries. Their true potential depends on effective alignment to ensure correct, trustworthy and ethical behavior, addressing challenges like misinformation, hallucinations, bias and misuse. While existing Reinforcement Learning (RL)-based alignment methods are notoriously complex, direct optimization approaches offer a simpler alternative. In this work, we introduce a novel direct optimization approach for LLM alignment by drawing on established Information Retrieval (IR) principles. We present a systematic framework that bridges LLM alignment and IR methodologies, mapping LLM generation and reward models to IR's retriever-reranker paradigm. Building on this foundation, we propose LLM Alignment as Retriever Preference Optimization (LarPO), a new alignment method that enhances overall alignment quality. Extensive experiments validate LarPO's effectiveness with 38.9 % and 13.7 % averaged improvement on AlpacaEval2 and MixEval-Hard respectively. Our work opens new avenues for advancing LLM alignment by integrating IR foundations, offering a promising direction for future research.