Hiding evidence from agents in a shared-model society improved compositional generalization by 20 points in 9 of 10 matched pairs

arXiv 2608.20054·low signalXBlueskyLinkedInCopy link

Four-cell agent societies sharing one frozen language model and LoRA adapter, communicating through continuous vectors, were run with and without full visibility of the input. Restricted-visibility arms beat their unrestricted twins by at least 20 percentage points in 9 of 10 matched pairs, held 0.558 on unseen composite functions, and same-value packet transplants preserved behavior at 0.94-1.00, implying a reusable value-indexed protocol emerged. The authors report the preregistered bar was missed (0.6988 median depth-three accuracy against a 0.70 floor), so treat this as a suggestive result on deliberately narrowing what each agent sees.