Apple Research Formalizes Theory of Evaluability for Black-Box Models

·2026.11.10 09:00

Key point

The study demonstrates that accurate evaluation of black-box models trained on test data is possible if and only if the data distribution is close to being small support.

Details

  • Formalizes a notion of evaluability for machine learning models without assuming the test set was held out from training, addressing challenges in evaluating large foundation models.
  • Fully characterizes the class of distributions where accurate black-box evaluation is possible, showing it is feasible if and only if the distribution is close to being small support.
  • Establishes tight upper and lower bounds on the sample complexity of evaluation in the natural setting where the evaluator only observes model outputs on the test set.
  • Demonstrates that a small amount of query access to the model can increase the evaluator's power for natural distributions and concept classes, connecting the problem to self-correctors in program testing.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.