Silent Dissent: LLM Agents That Yield to the Majority Still Represent Their Original Premise
4.20T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.02702.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryA pre-registered study shows that LLM agents in multi-agent debate who yield to a unanimous wrong majority still internally represent their original correct premise, detectable via a Jacobian lens on the residual stream. Stated consensus overstates actual agreement; hiding prior answers can raise conformity from 8% to 89%.
Why it mattersSurface consensus in LLM multi-agent debate can mask retained dissent. The J-lens method exposes this, relevant to any workflow that treats group agreement as a reliable signal.
Cited by
No citations on record.