Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.02827.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

Research paper showing that multi-agent LLM debates can amplify individual model biases into collective biased consensus, with LLM sampling temperature identified as a key driver. Proposes a physics-inspired phase-transition framework and shows that agent heterogeneity suppresses bias emergence, validated on investment and LLM-as-judge tasks.

Concrete design levers for multi-agent LLM systems: temperature and agent diversity measurably shift whether debate converges to biased consensus.

Cited by

No citations on record.