SPO: Multi-Dimensional Preference Sequential Alignment With Implicit Reward Modeling

TL;DR

SPO is a novel method for aligning large language models with multiple human preference dimensions by sequentially fine-tuning without explicit reward models, ensuring nuanced and multi-faceted alignment.

Contribution

The paper introduces SPO, a new approach that manages multi-dimensional human preferences in LLMs through sequential fine-tuning and theoretical analysis, avoiding explicit reward modeling.

Findings

SPO effectively aligns LLMs across multiple preference dimensions.

SPO outperforms baseline methods on various datasets.

Theoretical derivation of optimal SPO policy and loss function.

Abstract

Human preference alignment is critical in building powerful and reliable large language models (LLMs). However, current methods either ignore the multi-dimensionality of human preferences (e.g. helpfulness and harmlessness) or struggle with the complexity of managing multiple reward models. To address these issues, we propose Sequential Preference Optimization (SPO), a method that sequentially fine-tunes LLMs to align with multiple dimensions of human preferences. SPO avoids explicit reward modeling, directly optimizing the models to align with nuanced human preferences. We theoretically derive closed-form optimal SPO policy and loss function. Gradient analysis is conducted to show how SPO manages to fine-tune the LLMs while maintaining alignment on previously optimized dimensions. Empirical results on LLMs of different size and multiple evaluation datasets demonstrate that SPO successfully aligns LLMs across multiple dimensions of human preferences and significantly outperforms the baselines.