Calvin Isley (@calvinisley.com)

Calvin Isley

@calvinisley.com

Public Policy PhD Candidate @Harvard

Making my Bluesky debut with a few long-overdue paper announcements, beginning with my recent working paper on social sycophancy in LLMs. arxiv.org/abs/2609.26579 (with Johann Gaebler, @mlamparth.bsky.social, @juliaminson.bsky.social, and @5harad.com)

arxiv.orgReceptiveness, Not Sycophancy: Distinguishing Engagement from Deference in Language ModelsA central concern with language models is sycophancy: their tendency to defer to users' views at the expense of independent substantive judgment. In parallel, work on social sycophancy has focused on ...

Calvin Isley

@calvinisley.com

We find that measures of social sycophancy are entangled with desirable behaviors shown to help people engage constructively despite disagreement. Adding indicators of these behaviors to responses increases measured social sycophancy, even when the response’s substantive conclusion stays the same.

Calvin Isley

@calvinisley.com

Additionally, people prefer these responses and expect the advice recipient to be more willing to consider an opposing view. Training models to score lower on these measures of social sycophancy could therefore discourage constructive disagreement—something I’d like to see more of from LLMs!

You are up to date.