Published May 5, 2026 | Version v1

Misalignment or competent simulation: a reinterpretation of emotion vectors in large language models

Authors/Creators

Description

Recent work has identified internal patterns in large language models that correspond to emotional concepts and can be measured and manipulated. In “Emotion Concepts and Their Function in a Large Language Model,” these patterns, referred to as emotion vectors, are interpreted as functionally significant for understanding model behavior and alignment.

In this article, we re-examine these findings. We show that behaviors described as misaligned can be explained without introducing new internal constructs such as “functional emotions.” Across multiple case studies, the observed responses follow directly from the model acting as a highly capable simulator within the given context and constraints.

This leads to a central tension: a system trained to generate human-like responses cannot be expected to do so while simultaneously avoiding those same patterns of response. Behaviors classified as misaligned in these evaluations often resemble plausible human responses under similar conditions.

Interpreting these behaviors as misalignment introduces instability at the level of explanation, with direct implications for alignment research and the understanding of language models. Interventions aimed at modifying internal states risk targeting effects rather than causes if the underlying mechanism is context-sensitive simulation.

We conclude that these findings are consistent with large language models being very good at simulating human emotional patterns, rather than exhibiting functionally analogous emotional states.

Files

Misalignment or competent_anonymous.pdf

Files (236.0 kB)

Name Size
md5:8484f2c8ebbd966638bc34c98efc3205
236.0 kB Preview Download