Bottom line: Do LLMs have genuine preferences, or are they just mimicking human text? Evidence that something deeper than pattern-matching is occurring.
Examines whether LLMs show genuine preference coherence rather than mere mimicry. If the model's preferences are consistent across contexts, sensitive to trade-offs, and stable under reasonable prompt variation, that suggests something more like genuine preference than statistical pattern-matching. The paper finds evidence for coherence beyond what simple mimicry would predict.
If you ask someone what they want in one context, and then ask in a different way, a real preference shows up consistently. A mimic just says whatever sounds right in the moment. This paper tests whether AI models have consistent preferences or just sound good. The results suggest it's more than mimicry.