A new arXiv preprint, announced as arXiv:2610.24332, introduces a framework for PAC learning in which training examples do not reveal full labels. Instead, each example provides a single binary response to a query: whether the true label belongs to a chosen subset from a fixed query family. The paper focuses on the realizable multiclass setting, where a consistent hypothesis is assumed to exist.

The setup generalizes standard label supervision by allowing the learner to choose subsets adaptively, but the cost is that each example yields only one bit of information. This raises natural questions about sample complexity and learnability under restricted feedback.

However, the abstract as provided cuts off after the phrase "Our main…", so the paper's actual contributions and results are not stated in this source. Readers interested in the findings will need to consult the full manuscript.