dorsal/arxiv
View SchemaThe Straight and Narrow: Do LLMs Possess an Internal Moral Path?
| Authors | Luoming Hu, Jingjie Zeng, Liang Yang, Hongfei Lin |
|---|---|
| Categories | |
| ArXiv ID | 2601.10307vv1 |
| URL | https://arxiv.org/abs/2601.10307 |
| License | http://creativecommons.org/licenses/by-nc-sa/4.0/ |
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.
{
"annotation_id": "9884256c-5a1b-4e09-9408-9a0ca75520a2",
"date_created": "2026-02-17T05:53:23.836000Z",
"date_modified": "2026-02-17T05:53:23.836000Z",
"file_hash": "61e17028fe055be46087bcb3390e967b3618e31997713247ac9df7b5863ff32a",
"private": false,
"record": {
"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.",
"arxiv_id": "2601.10307",
"authors": [
"Luoming Hu",
"Jingjie Zeng",
"Liang Yang",
"Hongfei Lin"
],
"categories": [
"cs.CL"
],
"license": "http://creativecommons.org/licenses/by-nc-sa/4.0/",
"title": "The Straight and Narrow: Do LLMs Possess an Internal Moral Path?",
"url": "https://arxiv.org/abs/2601.10307",
"version": "v1"
},
"schema_id": "dorsal/arxiv",
"source": {
"execution_id": "e1854747-8077-4bb0-98f7-396145353ee1",
"id": "arXiv Dataset",
"type": "Model",
"variant": "snapshot-2026-01-17",
"version": "0.1.0"
},
"user_id": 1000002
}