Beyond Answer Sharing: TrueMAD for Targeted Peer Engagement in Multi-Agent Debate

Key points

• However, standard broadcast-style MAD makes peer answers visible without requiring agents to engage with a particular peer claim.1
• We propose TrueMAD, a backbone-agnostic test-time plug-in that introduces one-to-one, free-form peer advice between answer broadcasting and self-revision.1
• Each source agent responds to a target's current reasoning through support, rebuttal, refinement, or further questions, while the target retains control over how to revise.1
• Across AIME, GPQA-Diamond, and TruthfulQA with Qwen3-30B-A3B, TrueMAD improves agent-level and strict-majority accuracy over MAD by 3.62 and 3.52 percentage points on average, while TrueDMAD improves DMAD by 4.72 and 4.06 points, respectively.1

Sources (1)

• [1]Beyond Answer Sharing: TrueMAD for Targeted Peer Engagement in Multi-Agent Debate
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 12:08 PM

  Multi-agent debate (MAD) aims to improve large language model reasoning by exposing agents to alternative solutions and enabling iterative revision.

  However, standard broadcast-style MAD makes peer answers visible without requiring agents to engage with a particular peer claim.

Extractive summary: sentences quoted from the sources.