Abstract Priorities, Concrete Rewards: How Stated Preferences Relate to What People Reward in AI Responses
4.20T1 sourcearXiv cs.HC
Source record
Published by arXiv cs.HC (T1 source). The original is at https://arxiv.org/abs/2610.30763.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryAnalysis of 1,396 participants rating 29,347 LLM responses in the PRISM dataset finds that unasked-about features (length, formatting, refusal language) predicted ratings as strongly as surveyed attributes. Stated priorities tracked perceived model characteristics but not concrete response features; personalization from stated preferences added only 0.13 percentage points over pooled models.
Why it mattersQuantifies the gap between stated and revealed preferences in AI evaluation using a large public dataset. Challenges the premise that preference surveys meaningfully improve LLM personalization. Directly relevant to agent and tool builders.
Cited by
No citations on record.