Are the Values of LLMs Structurally Aligned with Humans? A Causal Perspective

TL;DR

This paper explores the underlying causal structure of LLMs' values, revealing differences from human values, and introduces lightweight, effective methods for more precise value alignment and steering.

Contribution

It proposes the concept of a causal value graph for LLMs and develops two novel, resource-efficient value-steering techniques based on this framework.

Findings

Causal value graphs differ significantly from human value systems.

Role-based prompting and SAE steering improve value alignment.

Experiments show enhanced control and effectiveness in LLMs.

Abstract

As large language models (LLMs) become increasingly integrated into critical applications, aligning their behavior with human values presents significant challenges. Current methods, such as Reinforcement Learning from Human Feedback (RLHF), typically focus on a limited set of coarse-grained values and are resource-intensive. Moreover, the correlations between these values remain implicit, leading to unclear explanations for value-steering outcomes. Our work argues that a latent causal value graph underlies the value dimensions of LLMs and that, despite alignment training, this structure remains significantly different from human value systems. We leverage these causal value graphs to guide two lightweight value-steering methods: role-based prompting and sparse autoencoder (SAE) steering, effectively mitigating unexpected side effects. Furthermore, SAE provides a more fine-grained approach to value steering. Experiments on Gemma-2B-IT and Llama3-8B-IT demonstrate the effectiveness and controllability of our methods.