Language-model groups overstate consensus when replaying human deliberation on a reasoning task

TL;DR - Replayed human reasoning discussions show that belief-anchored LLM agent groups substantially overstate consensus. Their agreement also fails to predict accuracy, with reasoning-mode agents sometimes converging almost unanimously on incorrect answers.

• Human full-consensus estimates varied from 24% to 57% depending on participation and final-state definitions.
• After submit-based and participation-matched adjustments, LLM groups still exceeded human consensus by roughly 34–44 percentage points.
• The gap persisted without early stopping and after removing the potentially memorizable answer.
• Simulated agent consensus did not match collective accuracy or reliably estimate human group outcomes.