# What You Can't See Is Still What You Learn: A Preregistered Sixty-Society Confirmation That Evidence Masking Drives Compositional Generalization

arXiv: https://arxiv.org/abs/2609.17637 · HTML: https://mlsanity.com/paper/2609.17637/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 5 of its 8 references that are in the map.
- First posted on arXiv 15 Sep 2026.

## Builds on (most cited first, 5 of 5)

- [Causal Abstractions of Neural Networks](https://mlsanity.com/paper/2106.02997/) (2021)
- [Compositionality and Generalization in Emergent Languages](https://mlsanity.com/paper/2004.09124/) (2020)
- [Natural Language Does Not Emerge 'Naturally' in Multi-Agent Dialog](https://mlsanity.com/paper/1706.08502/) (2017)
- [What You Can't See Is What You Learn: Slot-Selective Evidence Masking Favors Compositional Generalization in Shared-Genome Language-Model Societies](https://mlsanity.com/paper/2608.20054/) (2026)
- [Portable Semantics, Private Dialects: Reuse and Negative Transfer in Latent Communication Between Language-Model Cells](https://mlsanity.com/paper/2609.11365/) (2026)

## Abstract (arXiv)

Restricting what a module can read may improve what a system learns to compute. We test this in a preregistered confirmation with sixty four-cell systems sharing a frozen language-model backbone and communicating through learned continuous packets. Five conditions vary evidence masking, ownership markers, and replacement of foreign evidence with neutral filler (task-irrelevant text of the same token length), across six initialization clusters, each with two data orders, on one fresh task world. With markers available in both regimes, masking improved accuracy on held-out two- and three-operation compositions by median paired differences of 0.846 and 0.859; all twelve pairs cleared the required margins, and the full preregistered behavioral criterion passed. The unmarked replication also passed. No marked global-visibility (G+) system passed the marker-following check, so the effect of usable role information remains unresolved. The filler condition yielded seven full generalizers, but its decomposition criteria were inconclusive. Packet interventions in all eighteen audited masked systems followed the predicted intermediate-value changes on eligible cases; these finite, success-conditioned audits do not establish mediation. The results confirm a large advantage of the tested masking regime, while leaving its finer attribution and generality open. Protocols, results, and checkpoints are public.