Measuring debate shifts in AI model answers on multiple-choice tests
Measuring Collapse and Correction in Homogeneous-Panel LLM Debate
Computation and Language
Summary
This paper looks at how AI models argue over multiple-choice questions and whether their discussions really help or hurt the final answer quality. The researchers find that just seeing the final result improves doesn't tell the full story, because discussions can either fix mistakes or cause new errors. They develop a detailed way to track these changes during debates, showing when answers collapse or get corrected. Their tools help figure out when debate helps and when it might backfire.
What this means in practice
- •For ai system developers: Design evaluation protocols that distinguish when model debates improve or degrade multi-choice question answers.
- •For machine learning platform operators: Use pre-debate probes to triage risk of harmful answer collapses and better manage AI debate rounds.
Authors
Xin Li, Mengbing Liu, Chau Yuen
Abstract
Multi-agent large language model (LLM) debate is often evaluated by whether final answers improve, but movement is not necessarily improvement: the same discussion can rescue an initially wrong majority or destroy an initially correct one. Standard final-accuracy evaluations conflate these opposing mechanisms. We introduce an auditable protocol for homogeneous debate on multiple-choice questions (MCQs) that records each run as a transition ledger over collapse, correction, onset, and signed intervention utility. On 6,925 MMLU-Pro debates, the protocol identifies 253 collapses and a parallel correction ledger that changes how interventions should be judged. Replay experiments reveal the central tradeoff: a leave-one-model-out probe-gated freeze prevents 29 collapses but loses 108 corrections under equal weights, so collapse prevention alone can recommend the wrong policy. A compact pre-debate 8-probe screen is a triage signal: its unadjusted family-level association with conditional-collapse risk is high (G=7, Spearman rho=0.893, exact two-sided p=0.0123), but initial-majority accuracy is a close comparator (rho=0.821; family partial rho=0.767, p=0.0877), so we do not treat it as calibrated or capability-adjusted prediction. Round-level traces localize many collapses to the first debate round, where early disagreement can precede both harmful cascades and useful recovery. We release replayable schemas, coders, audits, cost cards, and zero-API rebuild scripts so future model-scaffold rows can be compared under the same denominators and signed utility ledger.