The Straight and Narrow: Do LLMs Possess an Internal Moral Path?

TL;DR

This paper explores the intrinsic moral representations of Large Language Models using Moral Foundations Theory, proposing a novel method to enhance their safety and alignment through internal moral vector manipulation.

Contribution

It introduces a cross-lingual probing technique, extracts steerable moral vectors, and develops Adaptive Moral Fusion for improved LLM safety and alignment.

Findings

Shared moral subspace across languages

Effective reduction of unsafe responses

Improved safety without sacrificing helpfulness

Abstract

Enhancing the moral alignment of Large Language Models (LLMs) is a critical challenge in AI safety. Current alignment techniques often act as superficial guardrails, leaving the intrinsic moral representations of LLMs largely untouched. In this paper, we bridge this gap by leveraging Moral Foundations Theory (MFT) to map and manipulate the fine-grained moral landscape of LLMs. Through cross-lingual linear probing, we validate the shared nature of moral representations in middle layers and uncover a shared yet different moral subspace between English and Chinese. Building upon this, we extract steerable Moral Vectors and successfully validate their efficacy at both internal and behavioral levels. Leveraging the high generalizability of morality, we propose Adaptive Moral Fusion (AMF), a dynamic inference-time intervention that synergizes probe detection with vector injection to tackle the safety-helpfulness trade-off. Empirical results confirm that our approach acts as a targeted intrinsic defense, effectively reducing incorrect refusals on benign queries while minimizing jailbreak success rates compared to standard baselines.