Who Convinced Whom? Tracing How LLMs Change Their Minds in Multi-Agent Systems
4.40T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.35930.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryThe study introduces MAD Bench, a 2,400-question benchmark for testing multi-agent debate, and evaluates three open-weight models against independent reflection. Results indicate that debate can produce small accuracy gains while also increasing shared errors, consensus, and sensitivity to peer identity and stated track record.
Why it mattersUseful for evaluators designing agent debates because it distinguishes small accuracy gains from increased conformity, unanimous errors, and source-status effects.
Cited by
No citations on record.