Confidence Composition for Multiagent Language Model Systems.

arXiv:2606.13591v2 Announce Type: replace Abstract: Multiagent language model systems, such as collaborative reasoning and debate, produce multiple correlated candidate answers and confidence signals. However, these signals are usually calibrated only at the i