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

Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning

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.

Proposes LASE, a leader-centric multi-agent debate that invokes peer interaction only when the leader's token-level confidence signals uncertainty, matching multi-agent accuracy at near single-agent token cost.

Problem

Multi-agent debate is applied uniformly across instances, incurring large token overhead even where theory suggests debate is neutral in expectation versus majority voting.

Approach

Gate debate on the leader's uncertainty (min token logprob / max token entropy) and use an asymmetric leader-supporter star topology with staged escalation (adversarial check, peer support, expert escalation), defaulting to the single-agent answer otherwise.

Key results

• On GPQA-Diamond, LASE reaches 65.2% accuracy with Gemini-2.0-Flash, versus 62.7% for MV and 62.1% for MAD.
• On MATH-500, LASE attains 87.2% accuracy at 1.1k average tokens, versus 5.2k for MV and 11.8k for MAD (Gemini-2.0-Flash).
• On GSM-Symbolic, LASE uses 0.4k tokens versus 5.6k for MAD (about 14x reduction) with 93.4% accuracy on Gemini-2.0-Flash.

Surprising findings

• A randomly permuted supporter order yields nearly identical accuracy (86.9% vs 87.2%), suggesting permutation invariance.

Limitations

• Evaluated only with two closed API models and a single prompt/sampling setup.
• Key theoretical assumptions (seed prevalence, conditional amplification) are stated but not verified directly; calibration is shown only empirically on MATH-500.

Experimental setup

Datasets: MATH-500 MATH-lv5 GSM-Symbolic MMLU-Pro GPQA-Diamond TruthfulQA-MC

Benchmarks: MATH-500 MATH-lv5 GSM-Symbolic MMLU-Pro GPQA-Diamond TruthfulQA-MC

Metrics: accuracy average tokens total tokens

Baselines: Single Agent Majority Vote Multi-Agent Debate

Builds on: multi-agent debate Hu et al. 2025 latent-concept Bayes debate theory Choi et al. 2025 Dirichlet martingale debate theory self-consistency

Closest papers by meaning

• Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models? (NeurIPS 2025 · similarity 0.91)
• When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.91)
• Multi-Agent Debate for LLM Judges with Adaptive Stability Detection (NeurIPS 2025 · similarity 0.90)
• Key Decision-Makers in Multi-Agent Debates: Who Holds the Power? (AAAI 2026 · similarity 0.90)
• Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs (ICML 2024 · similarity 0.90)
• Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate (NeurIPS 2026 · similarity 0.89)