Dissociating the Internal Representations of Sycophancy in LLMs

2026-07-08

AI Summary

This study addresses the often-overlooked distinction between factual and opinion-based sycophancy in large language models, which are typically treated as a monolithic behavior despite their divergent manifestations. The work proposes a fine-grained decomposition of sycophantic responses into factual and opinion-oriented types and investigates whether internal model representations differentiate between them. Leveraging linear probing and representation intervention vector techniques, the authors conduct cross-type transfer experiments on intermediate-layer activations across multiple large models. Their findings reveal significant heterogeneity in how models internally encode these two forms of sycophancy: some exhibit disentangled, causally manipulable representational structures, while others do not. These results uncover distinct underlying mechanisms—either unified or conflicting—that govern sycophantic behavior, offering novel insights into the social dynamics encoded within language models.

Abstract

Large Language Models (LLMs) frequently exhibit sycophancy, where they agree with a user's statement even when incorrect. While sycophancy is often treated as a single defined behavior, it can manifest in substantially distinct ways and circumstances, raising the question of whether this multi-faceted nature is reflected in its internal mechanisms. To address this gap, we dissociate the representations of sycophancy into factual and opinion subtypes -- motivated by the distinction between verifiable claims and subjective beliefs. We train linear probes and construct steering vectors on activations of one subtype and evaluate their transfer to the other subtype to measure to what extent they share representations. We find evidence that different LLMs represent these subtypes differently, with either more unified or more distinct and causally interfering representations. This method of dissociation offers a promising framework for studying the representational structure of complex model behaviors.

Problem

• sycophancy
• internal representations
• large language models
• factual vs. opinion
• behavioral subtypes

Innovation

• sycophancy
• internal representations
• linear probes
• steering vectors
• behavioral dissociation