
Eduardo Claudio de Carvalho Guedes Moreno · Zenodo (CERN European Organization for Nuclear Research) 2026 · 2026
DOI: 10.5281/zenodo.24265238
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
When a language model changes its answer after seeing other models, has it contributed something new or merely agreed? Studies of debate among LLMs usually read this change as peer influence, and recent work finds more than a third of answers changing even without peers, which is read as model instability. We show that this baseline depends on what the model is shown about itself. In a preregistered experiment (OSF aw9vf and 5dpjr), five models from five providers answered 224 multiple-choice questions, 140 from MMLU-Pro and 84 from a new bank, under five matched conditions: blind, blind again, seeing their own previous contribution, also seeing the peers' positions, and seeing the peers' full contributions. The measure, with no judge, is changing the option without declaring a new ground. On the questions where the panel started divided, answering again blind changed 23% of the answers; seeing one's own reasoning reduced this to 2%. Seeing the peers' contributions brought the change back to 20%: positions alone pulled little (+3.6 points) and the peers' reasoning pulled much more (+14.0 points). All five confirmatory tests passed after Holm correction, and accuracy barely changed. In an exploratory arm that shows the model only the letter of its previous answer, as in the most common control, the change rose back to 19% and 10%. Reconsidering is not instability: it depends on what the model sees of itself, and this difference explains part of the change attributed to peers.
Last synced
Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations
No comments yet — start the discussion below.