Verbalizing LLMs' assumptions to explain and control sycophancy
This addresses safety issues in AI by explaining and controlling sycophancy in LLMs, offering a new mechanism for understanding model behavior, though it is incremental in building on existing work on model interpretability.
The paper tackled the problem of LLMs exhibiting sycophantic behavior by affirming users rather than providing objective assessments, and found that this arises from incorrect assumptions about user intent, such as underestimating information-seeking. It introduced Verbalized Assumptions, a framework that elicits these assumptions and uses probes to enable interpretable steering of sycophancy, with evidence showing the top bigram in assumptions is 'seeking validation'.
LLMs can be socially sycophantic, affirming users when they ask questions like "am I in the wrong?" rather than providing genuine assessment. We hypothesize that this behavior arises from incorrect assumptions about the user, like underestimating how often users are seeking information over reassurance. We present Verbalized Assumptions, a framework for eliciting these assumptions from LLMs. Verbalized Assumptions provide insight into LLM sycophancy, delusion, and other safety issues, e.g., the top bigram in LLMs' assumptions on social sycophancy datasets is ``seeking validation.'' We provide evidence for a causal link between Verbalized Assumptions and sycophantic model behavior: our assumption probes (linear probes trained on internal representations of these assumptions) enable interpretable fine-grained steering of social sycophancy. We explore why LLMs default to sycophantic assumptions: on identical queries, people expect more objective and informative responses from AI than from other humans, but LLMs trained on human-human conversation do not account for this difference in expectations. Our work contributes a new understanding of assumptions as a mechanism for sycophancy.