From the source
LG AI Research is presenting work at ICML 2026 on reliability-based adaptive debate, arguing that more debate between AI agents is not automatically better.
- The post compares unstructured multi-agent debate to a meeting where everyone speaks freely from the start without a clear direction or role, which tends to run long without reaching good conclusions.
- It says debate works better when one agent first raises an idea that is crucial to the problem, after which the others validate and develop it.
- The post frames this with the terms "Correct Seed Scenario" and "Neutral Scenario": a valid idea, or "Correct Seed", can be amplified through interaction, while equal agents exchanging information with no new signals may get nowhere.
- It notes that two recent studies on the theory of multi-agent debate reach seemingly opposite conclusions.






