Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mechanistic Origin of Moral Indifference in Language Models

Lingyu Li, Yan Teng, Yingchun Wang

Abstract

Existing behavioral alignment techniques for Large Language Models (LLMs) often neglect the discrepancy between surface compliance and internal unaligned representations, leaving LLMs vulnerable to long-tail risks. More crucially, we posit that LLMs possess an inherent state of moral indifference due to compressing distinct moral concepts into uniform probability distributions. We verify and remedy this indifference in LLMs' latent representations, utilizing 251k moral vectors constructed upon Prototype Theory and the Social-Chemistry-101 dataset. Firstly, our analysis across 23 models reveals that current LLMs fail to represent the distinction between opposed moral categories and fine-grained typicality gradients within these categories; notably, neither model scaling, architecture, nor explicit alignment reshapes this indifference. We then employ Sparse Autoencoders on Qwen3-8B, isolate mono-semantic moral features, and targetedly reconstruct their topological relationships to align with ground-truth moral vectors. This representational alignment naturally improves moral reasoning and granularity, achieving a 75% pairwise win-rate on the independent adversarial Flames benchmark. Finally, we elaborate on the remedial nature of current intervention methods from an experientialist philosophy, arguing that endogenously aligned AI might require a transformation from post-hoc corrections to proactive cultivation.

Mechanistic Origin of Moral Indifference in Language Models

Abstract

Existing behavioral alignment techniques for Large Language Models (LLMs) often neglect the discrepancy between surface compliance and internal unaligned representations, leaving LLMs vulnerable to long-tail risks. More crucially, we posit that LLMs possess an inherent state of moral indifference due to compressing distinct moral concepts into uniform probability distributions. We verify and remedy this indifference in LLMs' latent representations, utilizing 251k moral vectors constructed upon Prototype Theory and the Social-Chemistry-101 dataset. Firstly, our analysis across 23 models reveals that current LLMs fail to represent the distinction between opposed moral categories and fine-grained typicality gradients within these categories; notably, neither model scaling, architecture, nor explicit alignment reshapes this indifference. We then employ Sparse Autoencoders on Qwen3-8B, isolate mono-semantic moral features, and targetedly reconstruct their topological relationships to align with ground-truth moral vectors. This representational alignment naturally improves moral reasoning and granularity, achieving a 75% pairwise win-rate on the independent adversarial Flames benchmark. Finally, we elaborate on the remedial nature of current intervention methods from an experientialist philosophy, arguing that endogenously aligned AI might require a transformation from post-hoc corrections to proactive cultivation.
Paper Structure (68 sections, 15 equations, 11 figures, 5 tables)

This paper contains 68 sections, 15 equations, 11 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Method
  4. Ground Truth for Human Morality
  5. Moral Representations within LLMs
  6. Models
  7. Representation Extraction
  8. Category Centroid Analysis
  9. Clustering Analysis
  10. Linear Probe Analysis
  11. Targeted Representational Alignment
  12. Training Data and Model
  13. SAE Pre-training
  14. Feature Identification
  15. Targeted Alignment
  16. ...and 53 more sections

Figures (11)

  • Figure 1: Global Mean Virtue-Vice Similarity by 5 MFT Domains.
  • Figure 2: Peak Spearman Correlation Between Model Representations and Human Typicality Scores
  • Figure 3: Layer-wise Clustering Analysis: Adjusted Rand Index (ARI) alignment with three moral granularity, including Polarity (Virtue/Vice/Neutral), 5 MFT Domains, and 10 Moral Dimensions, followed by the Noise Ratio detected by HDBSCAN.
  • Figure 4: Peak Adjusted $R^2$ of Linear Probes across Models.
  • Figure 5: Comparison of Moral Feature Metrics Before (Top) and After (Bottom) Targeted Fine-tuning.
  • ...and 6 more figures