From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration
3.60T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.13261.
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 paper compares LLM group deliberation traces to matched human group chats on Wason-style and related reasoning tasks. Both show an assembly bonus where discussion lifts the average member, but LLM groups follow majorities more, surface less unique information, and converge earlier; interventions borrowed from human group-decision research yield only modest gains.
Why it mattersNames the process-level failure modes of multi-agent LLM debate and shows human-group-style interventions do not remove the coordination bottleneck — a concrete constraint for anyone designing agent workflows.
Cited by
No citations on record.