Sycophancy as compositions of Atomic Psychometric Traits

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Jain, Shreyans, Yost, Alexandra, Abdullah, Amirali
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866909754748567552
author Jain, Shreyans
Yost, Alexandra
Abdullah, Amirali
author_facet Jain, Shreyans
Yost, Alexandra
Abdullah, Amirali
contents Sycophancy is a key behavioral risk in LLMs, yet is often treated as an isolated failure mode that occurs via a single causal mechanism. We instead propose modeling it as geometric and causal compositions of psychometric traits such as emotionality, openness, and agreeableness - similar to factor decomposition in psychometrics. Using Contrastive Activation Addition (CAA), we map activation directions to these factors and study how different combinations may give rise to sycophancy (e.g., high extraversion combined with low conscientiousness). This perspective allows for interpretable and compositional vector-based interventions like addition, subtraction and projection; that may be used to mitigate safety-critical behaviors in LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2508_19316
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Sycophancy as compositions of Atomic Psychometric Traits
Jain, Shreyans
Yost, Alexandra
Abdullah, Amirali
Artificial Intelligence
Computation and Language
Machine Learning
I.2.7; I.2.4
Sycophancy is a key behavioral risk in LLMs, yet is often treated as an isolated failure mode that occurs via a single causal mechanism. We instead propose modeling it as geometric and causal compositions of psychometric traits such as emotionality, openness, and agreeableness - similar to factor decomposition in psychometrics. Using Contrastive Activation Addition (CAA), we map activation directions to these factors and study how different combinations may give rise to sycophancy (e.g., high extraversion combined with low conscientiousness). This perspective allows for interpretable and compositional vector-based interventions like addition, subtraction and projection; that may be used to mitigate safety-critical behaviors in LLMs.
title Sycophancy as compositions of Atomic Psychometric Traits
topic Artificial Intelligence
Computation and Language
Machine Learning
I.2.7; I.2.4
url https://arxiv.org/abs/2508.19316