Multi-agent debate (MAD) has emerged as a promising approach for improving the reasoning ability of large language models (LLMs). However, existing turn-taking schemes typically disclose a speaker’s entire utterance before other agents can respond, allowing erroneous premises to spread through the shared context and making early correction difficult. This study proposes an interruptible MAD framework that enables early error correction through sentence-level disclosure and urgency-based turn-taking under a shared public-token budget. Each non-speaking agent continuously generates an action plan including its current assessment, action choice, urgency, and supported answer. The next speaker is then selected dynamically from agents requesting to speak or interrupt, while silent turns are allowed when no intervention is necessary. By revealing only one sentence at a time and discarding undisclosed sentences after interruption, the proposed framework is designed to prevent misleading claims from expanding into long incorrect explanations. Under a controlled evaluation on 1,000 MMLU questions using three agents with conditioned initial states containing both correct and incorrect answers, the proposed framework achieves the highest final accuracy in both the two-incorrect-one-correct setting (49.5% vs. 37.2% and 43.7%) and the one-incorrect-two-correct setting (79.2% vs. 68.7% and 73.8%). Analysis of intermediate answers further show that interruptions improve listeners’ answers more often than they worsen them. These results suggest that fine-grained, interruptible turn-taking can suppress misinformation propagation and stabilize consensus formation under the evaluated setting.
Kimura et al. (2026) studied this question.