Source-linked AI summary
When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning
Zihao Zhao, Tunyu Zhang, Haizhou Shi, Yusong Zhao, Xinxi Zhang, Hao Wang
TL;DR
Debate often fails to beat majority voting, raising the question of when interaction helps. The paper models debate as a combination of proposal supply and verification-aware readout, and tests both through neural-thicket societies and controlled interventions. Coverage-based societies and truth-sensitive readout improve debate outcomes across the tested backbones, although headroom conversion varies by task and backbone.
Problem
Multi-agent debate often fails to outperform majority voting, leaving unclear when interaction helps beyond voting.
Source: §1 · p. 1Method
The paper formalizes recoverable headroom, models verification-aware readout with LVD, and constructs neural-thicket societies using coverage objectives.
Source: §1 · p. 2Results
Coverage-based neural-thicket construction increases complementary correct proposals, and LVD uses answer-specific evidence to improve readout; matched stochastic evaluations show gains over Vanilla Debate across both backbones.
Source: §7 · p. 10Takeaways & Limitations
Debate's practical value depends on proposal headroom, the backbone's ability to evaluate competing evidence, and inference cost.
Source: §7 · p. 10Interaction can discard useful alternatives or damage vote-correct cases, and headroom conversion remains task- and backbone-dependent.
Source: §7 · p. 10
Abstract
from arXiv · showhide
Multi-agent debate can improve reasoning, yet often fails to beat simple majority voting. We argue that successful debate requires two distinct mechanisms: proposal supply must surface a correct answer, and readout must identify that answer when voting misses it. We formalize the first requirement through recoverable headroom, which measures cases where a correct proposal is available but the majority answer is wrong. For the second, we develop Latent Verification Debate (LVD), an accounting model in which candidate proposals receive answer-specific verification evidence before final generation. Controlled fixed-proposal interventions estimate this latent effect in equivalent peer-support units and show that correct evidence changes answer probabilities and generated decisions while proposal supply remains fixed. To improve proposal supply, we construct societies from neural-thicket agents using labeled and label-free coverage objectives. Across two backbones and matched-budget reasoning benchmarks, coverage-selected societies increase complementary proposal supply and improve aggregate accuracy in repeated stochastic evaluations. Round-level controls further show that interaction provides gains beyond applying the same finalizer directly to the initial proposals. These results identify proposal coverage and truth-sensitive evidence use as complementary conditions for debate to outperform voting. Code is available at https://github.com/Wang-ML-Lab/when-debate-helps.
1 Introduction
Debate often fails to outperform majority voting, motivating the question of when interaction helps. The paper separates that problem into supplying a correct proposal and reading it out when majority vote misses it.
1 Introduction
- Debate can mix proposals without improving expected correctness when interaction is symmetric and truth-agnostic, so interaction alone is not a guarantee over voting.The paper uses this martingale-null analysis as a boundary case for practical debate.Source: §1 · p. 1
- Debate helps only when proposal supply surfaces a correct answer and readout selects it, especially when a correct minority proposal is missed by majority vote.The paper calls these missed-but-surfaced cases recoverable headroom.Source: §1 · pp. 1–2
- Latent Verification Debate models candidate-specific verification evidence in readout, while coverage objectives select neural-thicket agents with complementary proposal behavior.The paper uses labeled and label-free coverage objectives for society construction.Source: §1 · p. 2
- Matched-budget tests use fixed-proposal interventions, repeated stochastic evaluations, and round-level controls to assess readout, robustness, and interaction gains.The evaluations span reasoning and multi-discipline benchmarks on Qwen3-4B and OLMo3-7B.Source: §1 · p. 2
2 Setup and Motivating Observations
The section defines the proposal and debate setup, then uses a truth-agnostic martingale null and neural-thicket observations to motivate a supply-readout account. Under fixed proposal supply, debate readout improves over voting, while remaining proposal-hit headroom indicates that readout does not recover every surfaced correct answer.
2. Setup and Motivating Observations
- The setup treats each agent's round-0 response as an independent proposal and later rounds as debate updates based on visible prior responses.The agents produce rationales and normalized answers; the answer values support voting and proposal-hit diagnostics.Source: §2.1 · p. 3
2.1 Proposal Setup and the Martingale Null
- Proposal counts update a symmetric answer-belief state but do not encode whether an answer is correct; under the symmetric truth-agnostic null, interaction adds no positive expected correctness drift.The paper presents this as a boundary case, not a complete account of practical debate.Source: §2.1 · p. 3
- The motivating observations contrast null-like dynamics for Qwen2.5-3B-Instruct with persistent gains over vote for Qwen3-4B-Instruct under the same protocol.The paper treats this contrast as an empirical boundary between neutral mixing and dynamics with additional structure.Source: §2.1 · p. 3, Figure 2 · p. 4
- Neural-thicket agents are nearby specialists formed by perturbing pretrained weights, and the method replaces scalar top-K selection with coverage-based society construction.The paper uses these heterogeneous agents to separate proposal supply from final readout.Source: §2.1 · pp. 3–4
- With the selected society held fixed, debate readout improves over round-0 vote, but proposal-hit headroom remains and correlates with final lift over vote.The proposal-hit upper bound marks cases where a correct answer was surfaced, not an available readout.Source: §2.1 · p. 4, Figure 2 · p. 4
3 Headroom Recovery Theory
The section separates debate into proposal supply and readout, defining recoverable headroom as correct answers surfaced but missed by majority vote. It then models verification-aware readout and specifies when evidence can recover a correct minority.
3.1 Recoverable Headroom and Gain
- Recoverable headroom is the opportunity where the society surfaces a correct answer but majority vote misses it.The section uses round-0 answers as the initial proposal supply and vote baseline.Source: §3.1 · p. 4, §3 · p. 4
- With initial proposals fixed, gain over vote decomposes into recovered headroom and new correctness, minus damage to vote-correct cases.Debate helps when recovery plus new correctness exceeds damage.Source: §3.1 · p. 5
3.2 LVD as Verification-Aware Readout
- LVD augments proposal-count readout with answer-specific verification evidence, while proposal-only readout lets answers gain influence through frequency alone.The evidence is a theoretical answer-level abstraction, not an observed hidden LLM state.Source: §3.2 · p. 5
- LVD increases posterior mass on the correct answer only when its verified evidence gives that answer more mass than the proposal-only baseline.The section frames this as the readout-level counterpart to the symmetric, truth-agnostic null.Source: §3.2 · p. 6
- On a headroom instance, LVD ranks a correct minority above the incorrect proposal-count winner only when verification evidence overcomes the minority’s proposal-layer disadvantage.Gains still require recoverable headroom and enough recovery to outweigh damage to vote-correct cases.Source: §3.2 · p. 6
4 Theory-Guided System Design
The design section translates the supply-readout framework into coverage-based society construction. It selects agents for complementary proposal coverage, using labeled answer coverage or a label-free agreement proxy.
4 Theory-Guided System Design
- The design targets more proposal hits and more verified evidence for surfaced correct candidates, pairing coverage-based society construction with verification-aware readout.More proposal hits can improve vote accuracy and create vote-missed correct proposals for readout to recover.Source: §4 · p. 6
4.1 Coverage-Based Society Construction
- Society construction favors agents that cover complementary proposal hits rather than simply selecting the individually highest-scoring agents.Construction approximates the ideal supply target on a split, using labels or an agreement proxy.Source: §4.1 · p. 6, §4.2 · p. 6
4.2 Answer-Coverage and Support-Coverage Greedy
- The question-level coverage objective counts a question once any selected agent receives a high proxy score, and greedy selection approximates its cardinality-constrained maximum-coverage objective.The greedy procedure chooses each next agent by marginal coverage increase.Source: §4.2 · p. 6
- Table 1 compares accuracy under a unified computation budget, using vote20 for vote rows and round-3 majority states for debate rows.Its average is the unweighted mean across eight displayed columns.Source: Table 1 · p. 7
- AC-Greedy uses labeled answer coverage, whereas SC-Greedy is label-free and scores agents through agreement among clustered valid answers.SC-Greedy can still select complementary agents rather than only majority-answer agents.Source: §4.2 · p. 7
5 Experiments
The experiments test proposal supply and readout across reasoning benchmarks, using coverage-selected societies, fixed-proposal interventions, transfer tests, and round-level controls. Coverage selection improves aggregate comparisons, while controlled evidence shifts and interaction gains show that debate’s benefits depend on converting available proposals into correct answers.
Diverse Expert Supply for Debate Systems
- Support-Coverage Greedy has the best average in the label-free block, while Answer-Coverage Greedy leads the label-based block across backbones.The gains are clearest on MMLU-Redux subjects, even when the aggregate average is close.Source: §5 · p. 8
Stochastic Robustness
- +1.14 and +1.46 points on Qwen3-4B, and +0.96 and +2.01 points on OLMo3-7B: SC-Greedy and AC-Greedy improve the eight-column average over Vanilla Debate under matched stochastic decoding.The paired 95% confidence intervals measure generation-seed variability, not uncertainty from resampling test questions.Source: §5 · p. 8
Can Coverage-Based Selection Transfer Out of Domain?
- On transfer from AMC12 selection to MMLU-Redux, the stronger coverage construction varies by backbone: Answer-Coverage Greedy leads final accuracy on Qwen3-4B, while Support-Coverage Greedy leads both outcomes on OLMo-3-7B.For Qwen3-4B, Support-Coverage Greedy instead has the largest improvement over the initial vote.Source: §5 · p. 8, Table 2 · p. 8
Dynamics and Mechanism Analysis
- R2 and R3 readouts reach 78.34% and 78.52%, respectively, exceeding the R0 readout by 1.89 and 2.07 points; OLMo MATH500 retains 3.33–6.33 points of headroom but has negative final lift.The same vote-then-judge finalizer raises R0 accuracy only from 76.09% to 76.45%, and later rounds can reduce answer diversity as consensus grows.Source: §5 · p. 8, Figure 3 · p. 9
Controlled Fixed-Proposal Readout Probes
- With proposal sets held fixed, correct peer evidence has positive margin effects in all four backbone–dataset combinations, and its answer-stripped rationale beats matched wrong reasoning on the gold-margin measure.At d = 3, correct support raises accuracy from .325 to .655 averaged across the four combinations, isolating truth-sensitive readout from proposal supply.Source: §5 · p. 9, Table 3 · p. 10
6 Related Work
The related work situates debate against majority-vote baselines and truth-agnostic interaction theory, then connects verification-aware readout and coverage-based agent selection to neighboring approaches.
Debate, Vote, and Null Models
Verification and Judging
- LVD places answer-specific verification within debate by changing proposal pseudo-counts before final generation, connecting readout to verifiers, judging, critique, and reranking [15].Its probabilistic treatment draws on hierarchical Bayesian deep learning while locating latent structure at multi-agent readout [27].Source: §6 · p. 10
Neural Thickets and Coverage Selection
- Coverage selection draws on neural thickets of perturbed models and budgeted subset-selection objectives, selecting complementary specialists rather than ranking agents only by average accuracy.Coverage signals may use true correctness or label-free answer-cluster support.Source: §6 · p. 10
7 Conclusion
The paper concludes that debate helps when proposal supply and readout work together, but its practical value remains conditional. The appendices specify the figure diagnostics, evaluation setup, compute accounting, and asset-use boundaries.
7 Conclusion
- Debate helps when coverage-based construction increases complementary correct proposals and answer-specific evidence improves their readout, but headroom conversion remains task- and backbone-dependent.The conclusion cautions that debate is not generally better than voting and that interaction may discard useful alternatives or damage vote-correct cases.Source: §7 · p. 10
A.1 Implementation Details for Fig. 2
- Figure 2 diagnostics use round-level accuracy relative to R0 vote, proposal-hit versus vote/debate bars, and a proposal-headroom versus final-lift scatter across matched cells.The scatter reports Pearson r = 0.77 and Spearman ρ = 0.82 across 20 cells.Source: §A.1 · p. 13
A.2 Evaluation Details
- Main evaluation uses data-specific generation budgets and fixed decoding within matched comparisons, with verification prompts and a tie-break judge only when the final vote has no unique winner.Experiments are inference-only and use BF16; reported hardware includes A100 80 GB and RTX PRO 6000 Blackwell 96 GB GPUs.Source: §A.2 · pp. 13–14, §A.3 · p. 14
A.3 Compute Resources
- The reported experiments use an estimated 0.9k GPU-hours, with a plausible range of 0.8–1.0k; queue time and CPU-only analysis are excluded.The broader project’s preliminary sweeps, failed or resumed jobs, and unreported-backbone experiments are also excluded.Source: §A.3 · p. 14, Table 4 · p. 14
A.4 Existing-Asset Licenses
- Model and dataset assets are used only for research evaluation, and model weights and raw benchmark questions are not redistributed.Source: §A.4 · p. 14
A.5 Five-Run Stochastic Robustness
Matched stochastic runs test coverage-selected societies and round-level controls across the robustness, prompt, and conversion analyses. The results show aggregate gains over Vanilla Debate, while prompt effects vary and interaction gains depend on converting available proposal headroom.
A.5 Five-Run Stochastic Robustness
- Both coverage-selected methods outperform Vanilla Debate in aggregate accuracy on both backbones; all four comparisons have nominal p < .05.Five matched stochastic runs use a shared generation seed within each run; the intervals capture generation randomness, not test-question resampling.Source: §A.5 · p. 15
A.6 Dataset and Split Details
- The main call-aligned evaluation covers AMC12, MATH500, GPQA, and MMLU-Redux; the diagnostic dynamics also include OlympiadBench.The held-out test sets range from 300 to 739 questions for the four main benchmarks.Source: §A.6 · p. 16
A.7 Seed Profiling Details
- Qwen3-4B-Instruct uses a 500-seed legacy pool, while OLMo’s downstream experiments retain scales 5 × 10^-4, 10^-3, and 2 × 10^-3.The realized Qwen scale counts are 160, 174, and 166; OLMo samples 100 seeds at each listed scale.Source: §A.7 · p. 16, Table 7 · p. 16
A.8 Prompt Templates
- The verification prompt asks agents to identify the most important correctness-relevant agreement or disagreement, whereas the standard prompt asks them to compare solutions and revise if needed.Both prompt variants require a final answer in the task’s specified format.Source: §A.8 · p. 17
A.9 Controlled Readout Probe Details
- Answer-stripped rationale probes compare correct, wrong, and neutral contexts using gold-answer-versus-wrong-mass and help-over-harm margins.The help-over-harm measure tests whether the helpful shift from correct reasoning exceeds the reverse harmful shift from wrong reasoning.Source: §A.9 · p. 18
- The standard prompt retains 2.37 points more R5 proposal-hit coverage, while final accuracies differ by only 0.07 points overall.Effects vary across datasets and selection methods, so the aggregate comparison does not establish a uniform prompt advantage or directly estimate verification mass.Source: §A.9 · p. 19, Table 12 · p. 19
A.10 Dynamics and Conversion Controls
- Applying the same finalizer to saved states yields 78.34% and 78.52% accuracy at R2 and R3, respectively, 1.89 and 2.07 points above the R0 result.The trajectory indicates rapid interaction gains followed by a plateau, rather than showing that the final round is always optimal.Source: §A.10 · p. 20, Table 13 · p. 20
- On OLMo MATH500, all four societies retain 3.33–6.33 points of initial headroom yet finish with negative lifts ranging from −1.67 to −0.33 points.This localizes the observed failure to conversion rather than missing proposals.Source: §A.10 · p. 20, Table 14 · p. 20
A.11 Secondary Construction and Transfer Diagnostics
Secondary diagnostics examine transfer, complementary proposal coverage, construction sensitivity, and round-level supply-to-readout conversion. They show that construction can provide complementary or answer-identity-free proposals, while recovery and damage shape whether that supply improves final accuracy.
A.11 Secondary Construction and Transfer Diagnostics
- SC-Greedy recovers 5.33 rather than 4.00 points on both backbones; on Qwen it also starts with more headroom and incurs less damage than AC-Greedy.On Qwen, AC-Greedy begins with the higher vote and retains the best final accuracy, while the larger SC-Greedy lift reflects recovery and lower damage.Source: §A.11 · p. 21, Table 15 · p. 21
- On 300 held-out questions, AC-Greedy covers 283, SC-Greedy 281, and Top-4 278; both coverage objectives select less-overlapping first agent pairs than Top-4.The evidence supports transferable, task-level complementary proposal behavior, not fixed semantic roles for selected agents.Source: §A.11 · p. 21, Table 16 · p. 21
- Increasing construction-set size from 25 to 150 improves AC-Greedy by 1.16 points and SC-Greedy by 0.56 points, while coverage is effectively saturated over N = 50–500.The perturbation scale that works best remains task- and backbone-dependent.Source: §A.11 · pp. 21–22, Table 17 · p. 22
- 84.30 hit@4 without labels or answer identities puts Hidden-Prefix within 0.37 points of supervised AC-Greedy; against TopAcc, it improves R5 final accuracy by 1.60 points.This answer-string-free construction result is on a multiple-choice task and does not validate semantic equivalence classes or open-ended generation.Source: §A.11 · p. 22, Table 18 · p. 22
A.12 Complete Round-Level Statistics
- Across methods, recoverable headroom falls from 7–8 points at R0 to about one point at R5 as debate compresses the proposal set.Useful minority answers are often surfaced before interaction, and later rounds can reduce alternatives even as vote accuracy improves.Source: §A.12 · p. 23
- AC-Greedy has higher average R0 proposal-hit rate than SC-Greedy, 84.93% versus 83.88%, but SC-Greedy has the larger average lift, 2.48 versus 1.64 points.The reported difference reflects SC-Greedy’s lower starting vote and greater conversion of missed proposals.Source: §A.12 · p. 23
A.12.1 Average Recovery Diagnostics
- Recovery falls sharply after the first rounds as debate compresses proposals; damage also falls, while new correctness becomes small after convergence.Later rounds can stabilize the majority while removing minority alternatives the readout could recover.Source: §A.12.1 · pp. 23–24, Table 19 · p. 24
A.12.2 Five-Round New-Seed Debate Results
Dataset-level results connect final lifts to initial proposal supply and the available room for improvement. Construction methods also differ in whether they yield the best final accuracy or the largest lift from their own initial vote.
A.12.2 Five-Round New-Seed Debate Results
- GPQA and MMLU-Redux have proposal-hit rates well above initial vote accuracy and the largest debate lifts, while high initial votes on MATH500 and AMC12 leave less room to improve.Final gain is largest when recoverable headroom is available and the readout converts it without large damage.Source: §A.12.2 · p. 24, Table 20 · p. 25
- AC-Greedy Debate has the best average final accuracy, whereas SC-Greedy Debate has the largest average lift from its own initial vote.This supports label-free coverage as a practical supply constructor while leaving supervised construction as a useful reference.Source: §A.12.2 · p. 24, Table 20 · p. 25
B Proofs for Sec. 3
The proofs establish the gain decomposition, the verification-aware posterior predictive, and conditions for verification evidence to affect readout rankings. They also connect marginalized verification support back to the same LVD readout form.
B.1 Proof of Proposition 1
- Proposition 1 decomposes final gain into recovered headroom and new correctness, minus damage to cases where the vote was correct.The proof identifies recovered headroom as H(S)Rη(S), new correctness as an additional positive term, and damage as the negative term.Source: §B.1 · p. 25
B.2 Proof of Lemma 1
- Lemma 1 shows that the verification-weighted pseudo-likelihood preserves a Dirichlet posterior, whose posterior predictive is the posterior mean.This supplies the conjugate update used in Eqn. 9.Source: §B.2 · pp. 25–26
B.3 Proof of Lemma 2
- Lemma 2 shows that marginalizing a latent verification draw yields the LVD readout form with soft verification support.The draw adds nonnegative mass to its selected answer.Source: §B.3 · p. 26
B.4 Proof of Proposition 2
- Proposition 2 establishes a fixed-vector condition for positive drift and a corresponding expected-drift condition when verification support is stochastic.The supplied proof fragments state these two cases without specifying their full expressions.Source: §B.4 · pp. 26–27
B.5 Proof of Proposition 3
- Proposition 3 reduces whether LVD ranks the correct answer above a verifier-supported alternative to comparing their numerators, since the denominator is shared.The resulting ranking condition is Eqn. 11.Source: §B.5 · p. 27
C Limitations
The paper bounds its claims by LVD’s dependence on surfaced correct proposals, a label-free proxy tied to answer clustering, and test-time compute that requires budget alignment.
C Limitations
- LVD can recover only surfaced correct proposals and can damage vote-correct cases, so it is a headroom-recovery method rather than a general guarantee over voting.When no correct proposal is present, the readout has no headroom to recover; experiments therefore report damage rate alongside final accuracy.Source: C Limitations · p. 27
- The label-free construction proxy depends on answer clustering and is most natural when valid answers can be normalized and grouped.Open-ended tasks may require a different proxy and a different definition of proposal-hit coverage.Source: C Limitations · p. 27
- The Hidden-Prefix result supports answer-string-free construction in a multiple-choice experiment, but does not validate semantic equivalence or evaluation for open-ended outputs.Source: C Limitations · p. 27
- Because the experiments use extra test-time compute, budget alignment is part of the empirical claim, with budgets matched or explicitly stated for vote, plain debate, and LVD.Source: C Limitations · p. 27