Research map / Multi-agent systems and games / LLM agents

When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Argues debate beats majority vote only when correct minority proposals are supplied and a verification-aware readout recovers them; proposes coverage-based neural-thicket society construction (label-based and label-free) and the LVD readout model, validated on Qwen3-4B and OLMo-3-7B.

Problem

Multi-agent debate often fails to beat simple majority voting; it is unclear when interaction adds value beyond vote, given the martingale-null theory that debate without truth-directed signal cannot improve expected correctness.

Approach

Decomposes debate gains into proposal supply (whether a correct answer is surfaced, measured as recoverable headroom) and readout (whether verification-aware evidence lets the system recover a correct minority answer). Builds agent societies from neural-thicket perturbations via answer-coverage or support-coverage greedy selection, and models verification evidence as pseudo-counts added to proposa

Key results

• Five-run stochastic robustness: SC-Greedy and AC-Greedy beat Vanilla Debate by +1.14 and +1.46 points on Qwen3-4B and +0.96 and +2.01 points on OLMo3-7B (paired 95% CIs reported).
• Round-level control: applying the same finalizer to R0 yields 76.09% accuracy vs 78.34% at R2 and 78.52% at R3, so interaction adds gains beyond an immediate finalizer.
• Controlled fixed-proposal probes: at d=3 correct peer support raises accuracy from .325 to .655 averaged across four backbone-dataset combinations.

Surprising findings

• On OLMo MATH500, all four societies retain 3.33–6.33 points of proposal-hit headroom yet have negative final lift, indicating a readout conversion bottleneck.
• The standard (non-verification) debate prompt keeps more proposal-hit coverage after R2 but does not raise final accuracy (0.07-point overall difference).

Limitations

• Debate gains are task- and backbone-dependent; headroom conversion fails on some tasks (e.g., OLMo MATH500).
• Equivalent-support and LVD weights are operational abstractions, not literal internal counters.
• Stochastic robustness intervals capture generation-seed variability only, not test-question resampling; some CIs lie near zero.

Experimental setup

Datasets: AMC12 MATH500 GPQA MMLU-Redux OlympiadBench

Benchmarks: AMC12 MATH500 GPQA MMLU-Redux OlympiadBench

Metrics: accuracy proposal-hit rate lift over vote paired 95% CI

Baselines: Self-Consistency (SC@20) Vanilla Debate Neural Thicket Vote Top-4 Debate Neural Thicket Debate

Builds on: neural thickets Vanilla Debate self-consistency Dirichlet-categorical-multinomial accounting

Closest papers by meaning

• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.91)
• Multi-LLM Debate: Framework, Principals, and Interventions (NeurIPS 2024 · similarity 0.91)
• Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate (ICML 2026 · similarity 0.91)
• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.91)
• Multi-Agent Debate for LLM Judges with Adaptive Stability Detection (NeurIPS 2025 · similarity 0.90)
• Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate (NeurIPS 2026 · similarity 0.90)