When should a language model answer directly, sample and vote, or engage in multi-agent debate? Recent work shows voting often explains much of the gain attributed to debate, while selective-debate systems activate deliberation only on uncertain examples. We ask: under a matched ceiling on generated tokens (960 per example), how much per-example routing headroom exists, and how much is recoverable from cheap pre-deliberation signals? We evaluate greedy decoding, three-sample voting, and a two-agent critique-revise debate on MuSiQue and GSM8K using Llama 3.1 8B Instruct and Ministral 3 8B Instruct. On MuSiQue, an oracle selecting the correct protocol per example gains +14.0 and +13.7 pp over the best fixed one. The best fixed protocol is model- and dataset-dependent: each (model, dataset) cell has a different winner. This headroom is hard to recover from cheap ex-ante signals. A vote-entropy threshold is the only controller that directionally beats the best fixed protocol on both models (+1.3 and +1.7 pp), though individual paired-bootstrap CIs include zero. A joint analysis (meta-analysis +1.6 pp, p=0.125; Bayesian P(both>0)=0.59) is directionally consistent but not significant. Learned controllers (LR, GBT) do not outperform the threshold. The key finding is structural: vote entropy predicts where debate is safe, not where debate is needed. High entropy sharply reduces debate backfire, but 66% of debate-helpful examples (31/47) occur when voting is unanimous but wrong. A single-prompt self-critique probe on Llama flips the answer in 127/127 unanimous cases, yielding zero mutual information with the debate-helpful label; we cannot rule out a prompt-compliance artifact, but either interpretation disqualifies the probe as a router. Recovering the remaining headroom requires behavioral probes that avoid format-compliance confounds at the 8B scale.
Multi-Agent Debate (MAD) with Majority Voting is a dominant paradigm for improving LLM reasoning, yet its effectiveness rests on the Condorcet Jury Theorem's assumption of independent errors. Because contemporary LLMs share similar pretraining corpora, their errors are strongly correlated, causing the majority to systematically suppress correct minority opinions, a phenomenon we term Minority Truth. Through debates among three heterogeneous LLM agents on six benchmarks, we find that roughly one in four divergent cases has the minority holding the correct answer, yielding a 10-percentage-point theoretical recovery margin. We propose Minority Sentinel, a lightweight meta-classifier that extracts a multi-dimensional debate fingerprint from debate logs and trains a LightGBM model to decide when to overturn majority voting. Minority Sentinel achieves a stable Flip Precision of 81.2% with positive Net Gain across all six datasets and all 20 random seed trials, demonstrating that debate logs contain sufficient behavioral signals for a non-LLM classifier to reliably recover suppressed minorities without degrading system accuracy. The LLM-as-Judge baseline yields negative Net Gain despite higher recall, confirming that flip safety, not recovery volume, determines intervention value.
Multi-agent debate, in which several LLMs exchange arguments before producing an answer, is widely assumed to improve answer quality by surfacing genuine disagreement. That disagreement is hard to verify, and no single signal can settle it, so we organize the analysis around four questions: (A) does the debater say it disagrees; (B) does its reply text actually argue; (C) does the dissent survive once the tone instruction that produced it is removed; and (D) do the probabilities assigned to stance options change? We evaluate three-model committees on 50 open-ended GlobalOpinionQA questions under three debate tones, friendly (seek common ground), neutral, and hostile (stress-test every position). The answers differ. (A) Full agreement differs by 50.4 percentage points between the friendly and hostile endpoints, pooling replies across all three rounds. (B) The text argues too: an LLM evaluator reading the contribution and reply without the structured self-report or tone instruction confirms the pushback is real. (C) Removing the instruction produces more returns to agreement in our sample, but the primary question-level test is inconclusive. (D) In a separate open-weight study, adjusted movement toward the opposing side is detected in two of three models, relative to a filler-based reference. This measures contextual response probabilities, not lasting belief change. For final answers, debate brings no measurable quality gain: an evaluator that judges each pair in both answer orders scores the debated answer no better than the same committee's no-debate answer on all 299 pairs, a test that detects severe but misses moderate damage in our checks. Taken together, LLM debate readily changes what agents say, but we find much weaker evidence that it changes what they persistently endorse or improves the quality of the final answer.
Multi-agent debate systems are typically evaluated only on whether the final answer is correct, overlooking the quality of the intermediate reasoning that debate is designed to produce. This paper studies the relationship between three signals in multi-agent debate: token-level log-probability distributions over reasoning tokens, LLM-as-judge rubric scores assigned to those tokens, and final task accuracy. We examine whether internal confidence signals predict externally evaluated reasoning quality, and whether either signal aligns with task correctness, across three domains: rubric-based scoring, mathematical reasoning, and factual question answering. Our framework pairs a two-agent debate architecture -- a Constructor and an Auditor -- with an LLM-as-judge that scores each agent's reasoning along instruction following, justification quality, and evidence grounding, together with a critical-failure flag. Experiments in the rubric-scoring domain reveal a consistent four-phase confidence trajectory and a substantial role asymmetry: confidence aligns with judged reasoning quality roughly twice as strongly for the Constructor as for the Auditor, and confidence-based detection of critical reasoning failures is markedly more reliable for the Constructor (AUROC 0.804) than for the Auditor (0.634). These findings motivate the broader cross-domain investigation proposed in this paper.