LG AI Research Proposes LASE Framework for Efficient Multi-Agent Reasoning at ICML 2026
·2026.11.17 09:00
Key point
The framework uses confidence-guided adaptive debate to reduce token usage to near single-agent levels while maintaining or improving accuracy over standard multi-agent methods.
Details
- LG AI Research introduced LASE (Leader-Adaptive Structured Engagement) at ICML 2026, a framework that argues multi-agent debate effectiveness depends on timing and structure rather than agent count.
- LASE achieved 87.2% accuracy on MATH-500 using Gemini-2.0-Flash with only 1.1k average tokens, outperforming Majority Voting (86.4% accuracy, 5.2k tokens) and standard Multi-Agent Debate (74.8% accuracy, 11.8k tokens).
- The method employs a Leader-Adaptive structure where a single Leader initiates debate only when uncertainty is high, using Min Logprob and Max Entropy as confidence proxies to trigger intervention.
- Interaction follows a 1:1 Directed Star Topology where Supporters provide targeted signals (Adversarial Check, Peer Support, Expert Agent) to the Leader, avoiding symmetric all-to-all debates that may not improve expected accuracy.
- Calibration analysis showed that high Min Logprob (0.99+) correlates with 85.5% accuracy, while low confidence (below 0.50) drops to 50.9%, validating the uncertainty gate as a reliable error signal.
- On smaller models like Llama-3-8B, LASE achieved 32.9% accuracy on MATH-500, surpassing Single Agent (25.7%) and standard Multi-Agent Debate (24.1%), demonstrating benefits even with limited base model capabilities.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.