Blocking an AI module from seeing everything its neighbors know just made a batch of test systems dramatically better at combining skills they had never combined before.
Researchers ran sixty four-cell AI systems built on a shared, frozen language-model backbone, letting the cells talk only through compressed packets of information rather than raw data, according to a preregistered study posted September 17 on arXiv titled "What You Can't See Is Still What You Learn: A Preregistered Sixty-Society Confirmation That Evidence Masking Drives Compositional Generalization" (arXiv:2609.17637). Some systems let every cell see all the evidence; others masked parts of it, with or without labels marking who owned which piece. On tasks that required chaining two or three learned operations on unfamiliar inputs, the masked systems beat the fully visible ones by wide margins - median accuracy gains of roughly 0.85 across all twelve tested pairings, clearing every threshold the researchers set in advance. A version that swapped hidden evidence for meaningless filler still produced seven systems that generalized fully, though the researchers could not explain why.
The result feeds a growing argument in multi-agent AI design: giving every component the full picture is not obviously the safest default, and restricting information flow can force a system to learn general, reusable rules instead of memorizing shortcuts. It is also an unusually large, preregistered test for this field - sixty systems, six initialization clusters, two data orders each - in an area where single anecdote demos often pass for evidence.
The catch is attribution: the study's own internal audits could only confirm the predicted mechanism in systems that already succeeded, so nobody yet knows precisely what masking is teaching these systems to do differently - only that, in this setup, it works.