ArXiv 2609.17637: masking yields 0.846 median accuracy gain on held-out compositions in 60-system preregistered test

Sixty four-cell systems shared a frozen language-model backbone and exchanged learned continuous packets. Five conditions manipulated evidence masking, ownership markers, and neutral filler replacement. Masking regimes with markers passed the full preregistered behavioral criterion; unmarked replication also passed. Packet interventions in audited masked systems produced the predicted intermediate-value shifts on eligible cases. No globally visible system satisfied the marker-following check.

The study isolates masking from visibility by holding the backbone fixed and varying only packet access rules. Median gains exceed 0.84 with every pair clearing the margin, indicating the effect is not initialization- or order-dependent. Filler replacement produced seven full generalizers yet failed decomposition criteria, leaving mediation unresolved. These outcomes align with prior attention-masking results in transformers where restricted context improved out-of-distribution composition without altering parameter count.

Operationally the protocol supplies public checkpoints and audit logs that permit direct replication of packet states. Future training pipelines can insert equivalent read restrictions at negligible compute cost. The unresolved role of explicit markers suggests targeted follow-up experiments that decouple marker presence from masking depth. No globally visible baseline cleared the criterion, constraining claims about scale alone.

Subsequent work should test whether the same masking schedule transfers to decoder-only models above 7B parameters and whether the accuracy delta persists beyond three operations.