Multi-agent committees adopt a wrong answer 38% of the time when two peers assert it — and only an independent referee catches it

arXiv·medium signalXBlueskyLinkedInCopy link

Across seven cohorts on six public clinical datasets spanning text, imaging, and tabular ICU records, Gemini committees resisted isolated shortcut cues (5-16% flip) but were highly vulnerable to socially plausible ones: when two peers asserted the same wrong answer the holdout adopted it 38% of the time, as did a fabricated "pre-screen" system flag, on both capability tiers. Of three oversight designs, a gate could not separate adoption from honest agreement (100% false-positive rate) and a same-lineage transcript judge worked on text (100% precision, 93% recall) but collapsed on imaging; only a referee that privately re-queried the holdout transferred, at 77-88% precision. Tripling a cue's visual salience did nothing while a second peer voice raised contagion by half again — the lesson being that any multi-agent debate or self-consistency architecture needs an oversight channel independent of the agents' self-reports.