Sycophancy as compositions of Atomic Psychometric Traits
This work addresses sycophancy, a key behavioral risk in LLMs, by providing a novel compositional framework for interpretable interventions, though it appears incremental in its methodological approach.
The authors tackled the problem of sycophancy in LLMs by modeling it as compositions of psychometric traits, enabling interpretable vector-based interventions to mitigate this safety-critical behavior.
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.