arXiv cs.CL· Luoming Hu, Jingjie Zeng, Liang Yang, Hongfei Lin·· 14 小时前AI 评分38
研究：LLM 是否存在内在道德路径？基于道德基础理论的探测与干预
The Straight and Narrow: Do LLMs Possess an Internal Moral Path?
AI 导读
研究者利用道德基础理论（MFT）对 LLM 的道德表征进行跨语言线性探测，发现中英文在中层共享道德表征但存在不同的道德子空间，并据此提取出可操控的 Moral Vectors。基于此提出的 Adaptive Moral Fusion（AMF）在推理时结合探测与向量注入，相比标准基线在减少对良性查询的错误拒绝的同时降低了越狱成功率。该论文已被 EMNLP 2026 Main 接收。
正文 · 原文
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.
| Comments: | Our paper has been accepted by EMNLP 2026 Main; Code and datasets: this https URL |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2601.10307 [cs.CL] |
| (or arXiv:2601.10307v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2601.10307 arXiv-issued DOI via DataCite |
Submission history
From: Luoming Hu [view email]
[v1]
Thu, 15 Jan 2026 11:42:00 UTC (1,647 KB)
[v2]
Mon, 26 Oct 2026 17:26:51 UTC (1,518 KB)
来源：arXiv cs.CL · arxiv.org