Andrei Chetvergov

Models exhibit a surprising preference for their own prior answers, revealing inconsistencies in LLM value profiling across response formats.

Scalar favorability fails to capture the full spectrum of affective nuances, revealing the need for a more sophisticated profiling approach in LLM evaluations.

LLMs misidentify adjacent values over 50% of the time, revealing critical biases that could skew their understanding of human motivations.

Affective framing in LLMs reveals that models with templated responses cluster together at low emotional arousal, challenging traditional sentiment analysis approaches.