Metaphors are a Source of Cross-Domain Misalignment of Large Reasoning Models

2026-01-06

arXiv.org

AI Summary

This study investigates how metaphors induce alignment biases in large language models during cross-domain reasoning. The authors demonstrate that metaphors present in training data can disrupt the model’s reasoning pathways, leading to failures in cross-domain alignment. To address this, they establish, for the first time, a causal link between metaphor usage and alignment bias, and introduce an analytical framework based on metaphor intervention. Their approach integrates causal intervention techniques, latent variable feature activation analysis, and a novel detector designed to identify alignment biases in hidden-layer representations. Experimental results show that metaphor-based interventions effectively modulate cross-domain alignment behavior, and the proposed detector accurately predicts bias content with high precision. This work offers a new perspective for understanding and mitigating alignment risks in large language models.

Technology Category

• Natural Language Processing: Safety and Robustness
• Cognitive Modeling & Cognitive Systems: Analogy
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Earlier research has shown that metaphors influence human's decision making, which raises the question of whether metaphors also influence large language models (LLMs)'reasoning pathways, considering their training data contain a large number of metaphors. In this work, we investigate the problem in the scope of the emergent misalignment problem where LLMs can generalize patterns learned from misaligned content in one domain to another domain. We discover a strong causal relationship between metaphors in training data and the misalignment degree of LLMs'reasoning contents. With interventions using metaphors in pre-training, fine-tuning and re-alignment phases, models'cross-domain misalignment degrees change significantly. As we delve deeper into the causes behind this phenomenon, we observe that there is a connection between metaphors and the activation of global and local latent features of large reasoning models. By monitoring these latent features, we design a detector that predict misaligned content with high accuracy.

Problem

• metaphors
• cross-domain misalignment
• large language models
• reasoning pathways
• latent features

Innovation

• metaphor
• cross-domain misalignment
• large language models
• latent features
• causal intervention