Saved in:
Bibliographic Details
Main Authors: Wang, Mengdie Flora, Xie, Haochen, Wang, Guanghui, Gao, Aijing, Yang, Guang, Li, Ziyuan, Qiu, Qucy Wei, Han, Fangwei, Qiu, Hengzhi, Huang, Yajing, Zhu, Bing, Woo, Jae Oh
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.07667
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915926316679168
author Wang, Mengdie Flora
Xie, Haochen
Wang, Guanghui
Gao, Aijing
Yang, Guang
Li, Ziyuan
Qiu, Qucy Wei
Han, Fangwei
Qiu, Hengzhi
Huang, Yajing
Zhu, Bing
Woo, Jae Oh
author_facet Wang, Mengdie Flora
Xie, Haochen
Wang, Guanghui
Gao, Aijing
Yang, Guang
Li, Ziyuan
Qiu, Qucy Wei
Han, Fangwei
Qiu, Hengzhi
Huang, Yajing
Zhu, Bing
Woo, Jae Oh
contents Multi-agent debate improves LLM reasoning, yet agreement among agents is not evidence of correctness. When agents converge on a wrong answer through social reinforcement, consensus-based stopping commits that error to an automated action with no recourse. We introduce Conformal Social Choice, a post-hoc decision layer that converts debate outputs into calibrated act-versus-escalate decisions. Verbalized probability distributions from heterogeneous agents are aggregated via a linear opinion pool and calibrated with split conformal prediction, yielding prediction sets with a marginal coverage guarantee: the correct answer is included with probability ${\geq}\,1{-}α$, without assumptions on individual model calibration. A hierarchical action policy maps singleton sets to autonomous action and larger sets to human escalation. On eight MMLU-Pro domains with three agents (Claude Haiku, DeepSeek-R1, Qwen-3 32B), coverage stays within 1--2 points of the target. The key finding is not that debate becomes more accurate, but that the conformal layer makes its failures actionable: 81.9% of wrong-consensus cases are intercepted at $α{=}0.05$. Because the layer refuses to act on cases where debate is confidently wrong, the remaining conformal singletons reach 90.0--96.8% accuracy (up to 22.1pp above consensus stopping) -- a selection effect, not a reasoning improvement. This safety comes at the cost of automation, but the operating point is user-adjustable via $α$.
format Preprint
id arxiv_https___arxiv_org_abs_2604_07667
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle From Debate to Decision: Conformal Social Choice for Safe Multi-Agent Deliberation
Wang, Mengdie Flora
Xie, Haochen
Wang, Guanghui
Gao, Aijing
Yang, Guang
Li, Ziyuan
Qiu, Qucy Wei
Han, Fangwei
Qiu, Hengzhi
Huang, Yajing
Zhu, Bing
Woo, Jae Oh
Artificial Intelligence
Multiagent Systems
Social and Information Networks
Multi-agent debate improves LLM reasoning, yet agreement among agents is not evidence of correctness. When agents converge on a wrong answer through social reinforcement, consensus-based stopping commits that error to an automated action with no recourse. We introduce Conformal Social Choice, a post-hoc decision layer that converts debate outputs into calibrated act-versus-escalate decisions. Verbalized probability distributions from heterogeneous agents are aggregated via a linear opinion pool and calibrated with split conformal prediction, yielding prediction sets with a marginal coverage guarantee: the correct answer is included with probability ${\geq}\,1{-}α$, without assumptions on individual model calibration. A hierarchical action policy maps singleton sets to autonomous action and larger sets to human escalation. On eight MMLU-Pro domains with three agents (Claude Haiku, DeepSeek-R1, Qwen-3 32B), coverage stays within 1--2 points of the target. The key finding is not that debate becomes more accurate, but that the conformal layer makes its failures actionable: 81.9% of wrong-consensus cases are intercepted at $α{=}0.05$. Because the layer refuses to act on cases where debate is confidently wrong, the remaining conformal singletons reach 90.0--96.8% accuracy (up to 22.1pp above consensus stopping) -- a selection effect, not a reasoning improvement. This safety comes at the cost of automation, but the operating point is user-adjustable via $α$.
title From Debate to Decision: Conformal Social Choice for Safe Multi-Agent Deliberation
topic Artificial Intelligence
Multiagent Systems
Social and Information Networks
url https://arxiv.org/abs/2604.07667