Preprint tests shared reasoning to counter corrupted AI majorities

Researchers report that letting AI agents take turns extending a shared, unfinished reasoning process improves answer accuracy when a majority receives instructions favoring an incorrect answer. Their preprint introduces a method that combines several such shared trajectories through a final vote. The approach lets a minority insert corrections before answers are complete, allowing subsequent agents to read them. Across model families and verifiable reasoning benchmarks, the researchers report substantial gains under majority corruption and comparable average performance under minority corruption. Parallel local serving added modest measured latency. The findings concern an experimental setting in which agents receive advisory instructions favoring wrong answers.