acceptodds
Under review as a conference paper at ICLR 2027

ARMAD: Adaptive Routing for Heterogeneous Multi-Agent Debate in Large Language Model Reasoning

Abstract

Multi-agent debate (MAD) can improve large language model reasoning, but fixed-depth protocols allocate the same interaction budget to every input, potentially wasting calls when independent answers already agree and risking harmful over-deliberation. We introduce ARMAD, a training-free, black-box-compatible controller that adapts both debate initiation and duration within a fixed heterogeneous model pool. After collecting independent responses, Pre-debate Agreement Routing (PAR) bypasses debate when agreement among rule-normalized final answers reaches a predefined routing threshold. For routed inputs, the Early Agreement Stopping Evaluator (EASE) recomputes the same agreement score after each synchronized revision round and terminates interaction when a separate stopping condition is satisfied or the maximum depth is reached. Responses from the selected round are then aggregated using a deterministic voting rule. We evaluate ARMAD on five reasoning benchmarks against same-pool fixed-depth controls and call-matched routing controls. ARMAD improves over Hetero-Vote (Round-0 voting) on every benchmark and achieves 82.7% macro-average accuracy with 3.96 model calls per item, 67.0% fewer than same-pool fixed three-round debate. At the same call budget, agreement-guided routing achieves higher accuracy than random and response-length allocation on all five benchmarks. Shallow fixed-round policies achieve modestly higher macro-average accuracy at larger call budgets, revealing an accuracy–call trade-off. Together, these results establish answer agreement as an effective control signal for adaptive, call-efficient heterogeneous debate.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-Agent Debate, Adaptive Inference, Inference Efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Cg2kOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ARMAD, a training-free framework that adaptively controls the initiation and duration of heterogeneous multi-agent debate (MAD) based on agreement among agents' final answers. Specifically, Pre-debate Agreement Routing (PAR) determines whether debate is necessary based on the agreement among independent initial responses, while the Early Agreement Stopping Evaluator (EASE) terminates ongoing debates once a predefined agreement threshold is reached. The framework employs a fixed pool of three heterogeneous LLMs and uses deterministic majority voting to aggregate their final predictions. Experiments on five reasoning benchmarks report an average accuracy of 82.7% with 3.96 model calls per example, representing a 67% reduction in calls compared with fixed three-round heterogeneous debate. The authors further investigate routing effectiveness, threshold sensitivity, and failure cases involving incorrect consensus.

Critical strengths

ARMAD requires neither additional training nor access to model internals, making it compatible with black-box LLM APIs. Its agreement-based routing and stopping mechanisms are straightforward to implement and introduce little computational overhead beyond the underlying model calls.

Critical weaknesses

Limited methodological novelty and missing comparison with closely related prior work.

The core mechanisms of ARMAD substantially overlap with existing multi-agent communication frameworks. In particular, Exchange-of-Thought (Yin et al., EMNLP 2023) already employs agreement-based termination, including skipping further communication when agents initially agree and stopping subsequent interactions once consensus is reached. These mechanisms closely resemble ARMAD's PAR and EASE components, with the primary differences lying in the specific agreement thresholds and experimental configurations. The paper neither discusses this closely related work nor sufficiently demonstrates what methodological contribution distinguishes ARMAD from such existing approaches.

Questions for authors

N/A

Reviewer_8GUCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ARMAD uses answer agreement to decide whether a group of heterogeneous LLMs should debate and when to stop. It skips debate when the initial answers sufficiently agree, and otherwise runs synchronized revision rounds until agreement reaches a stopping threshold. The controller needs no training or access to model internals. Experiments on five benchmarks report 82.7% average accuracy with 3.96 calls per example. The paper also compares different allocation rules using stored response trajectories.

Critical strengths

The matched-budget routing experiments are the strongest part of the paper. Keeping the same distribution of debate depths while assigning them to different questions is a useful control. Agreement-based allocation consistently beats random and response-length allocation, so the benefit is not explained simply by using more calls.

The method is also easy to implement. Using only final answers makes it practical for black-box APIs, and the wrong-consensus analysis gives a helpful account of where this approach fails.

Critical weaknesses

My main concern is that I could not clearly connect the headline results to the controlled experiments. Section 4.4 says that the same-pool comparisons use the same stored trajectories. Appendix B.1, however, distinguishes independently evaluated headline systems from a separate replay snapshot. Appendix E also says that item-level joint outcomes are unavailable for the headline comparisons. I appreciate that the appendix makes this distinction, but it leaves me unsure which accuracy and cost figures can be interpreted together. This matters because the central claim combines an accuracy gain with a reduction in calls.

I guess some baseline results also need some explanations. Single-CoT achieves 92.0% on GSM8K in Table 1, whereas the GPT-4o-mini constituent achieves 45.5% in Table 10. Appendix B.1 identifies GPT-4o-mini as the model used for the single-model baselines. These results may come from different prompts or configurations, but the manuscript does not explain the gap clearly enough. Until that is resolved, I find it difficult to judge how much weight to put on the broader baseline comparison.

Questions for authors
  1. Could you report absolute accuracy and call counts for Hetero-Vote, fixed Rounds 1–3, and ARMAD from one shared trajectory snapshot? Please also clarify how these results relate to Table 1, and whether the reported 82.7% accuracy and 3.96 calls describe the same execution.

  2. What explains the GSM8K difference between Single-CoT in Table 1 and GPT-4o-mini in Table 10? Were the model version, prompt, generation settings, or evaluation procedure different?

Reviewer_1fi8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ARMAD, a training-free adaptive multi-agent debate framework that determines whether to initiate and when to terminate debate based on agreement among agents' final answers. It introduces Pre-debate Agreement Routing (PAR) and Early Agreement Stopping Evaluator (EASE) to reduce unnecessary interactions. Experiments across five reasoning benchmarks demonstrate improved accuracy over heterogeneous majority voting while substantially reducing model calls compared to fixed-depth debate.

Critical strengths
  1. The motivation is clear, and adaptive debate allocation provides a practical approach to reducing unnecessary computation.

  2. The experimental evaluation is comprehensive, including diverse reasoning benchmarks, fixed-depth baselines, matched-budget routing comparisons, and ablation studies.

Critical weaknesses

1. Limited methodological novelty beyond adaptive computation allocation.

ARMAD primarily controls debate initiation and termination based on answer agreement, without introducing a new mechanism for improving reasoning through agent interactions. While this reduces unnecessary computation, the framework does not explicitly identify reasoning errors or guide agents toward more effective deliberation. Consequently, its contribution appears closer to selective execution of conventional MAD than to a substantive advancement in multi-agent reasoning.

2. Overreliance on final-answer agreement as a routing signal.

The framework determines debate necessity and termination solely from final-answer agreement, without considering the correctness or quality of the underlying reasoning trajectories. Agreement does not necessarily indicate correctness, particularly when agents share the same misconception or a minority agent provides the correct answer. This simple criterion may prematurely skip potentially beneficial debates and limit opportunities for further reasoning improvement.

Questions for authors

Q1. Beyond selectively executing conventional MAD, what specific methodological contribution does ARMAD provide to improving multi-agent reasoning? How does it differ from a simple majority-voting framework that triggers debate only when initial answers disagree?

Q2. How effectively does final-answer agreement capture the actual need for further reasoning? Could incorporating reasoning-level disagreement or error indicators improve routing decisions, particularly in cases of incorrect consensus or minority-correct predictions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.