Dense linear probes are a common tool for assessing whether a language model's internal representations contain enough information to predict a given task. But they only measure readability, not how many distinct features are actually doing the work. A new paper from arXiv introduces a complementary metric, k*, which quantifies readout sparsity: the smallest number of features that still supports accurate prediction.
The authors argue that knowing k* gives a more direct picture of how a model organizes task information. A low k* would suggest the model relies on a compact set of features, while a high k* implies the task is spread across many dimensions. This distinction is invisible to standard dense probes, which assume all features are equally available.
The abstract is brief, and the full results are not yet available in the summary. Still, the proposal offers a practical way to move beyond simple readability and toward a finer-grained understanding of representational structure in language models.