Researchers have proven the effectiveness of data masking for training AI

A group of scientists published research results, confirming that limiting the volume of data available for processing by individual system modules significantly improves their ability to compositional generalization. During the experiment, 60 systems were used, built on the basis of a frozen language model, which exchanged information through trainable continuous packets.

As part of the testing, various conditions for masking evidence were applied, using possession markers and replacing external data with neutral fillers. With the presence of markers, data masking increased the accuracy of task execution by 0.846 and 0.859 median scores for compositions of two and three operations, respectively. All 12 control pairs successfully passed the established accuracy thresholds, confirming the hypothesis about the benefits of limited access to information.

Researchers conducted an audit of 18 systems with masking applied, in each of which they recorded predicted changes in intermediate values when processing data. Despite success in tests, the impact of explicit role markers on system behavior remained not fully studied, as none of the globally visible models passed the test for following markers.

The authors of the work provided open access to protocols, results, and model checkpoints in accordance with the arXiv license. The obtained data indicate that architectural restrictions on data visibility can become a key factor in the development of more reliable and logically consistent AI systems that can better cope with new combinations of tasks.