Apple AI Updates: July 18, 2026

1. Apple Shows Machine Unlearning Can Skip Low-Influence Data Points

Apple. Apple researchers challenged the common assumption that every data point must be actively removed during machine unlearning, arguing that points with negligible impact on model outputs can be filtered out first. Using influence functions to identify low-influence training data across language and vision tasks, the team reduced the computational cost of unlearning by up to roughly 50 percent in empirical tests. The work aims to make privacy-preserving data removal more practical to deploy without weakening privacy protections. Source

2. Apple Introduces VICIS for Inferring Visual Concepts From Image Sets

Apple. Apple presented VICIS (Visual Concept Inference from Sets), a task and model in which a system is shown a small set of images sharing a concept plus a query image, then must generate new images that preserve the concept while matching the query. The researchers reported that state-of-the-art vision-language models perform poorly on this setup, often ignoring visual context or producing biased outputs. Their proposed architecture extracts concept-specific embeddings and generated more accurate and diverse results, generalizing to unseen concepts and other modalities such as sketches, with evaluation on synthetic data and large-scale ImageNet and WordNet sets. Source

3. Apple Extends Interactive Proofs to General Distribution Properties

Apple. Apple researchers built interactive proof systems that let a verifier with limited samples validate a prover’s claims about an unknown distribution without repeating the expensive analysis. The approach covers properties decidable by bounded-depth circuits or Turing machines and achieves doubly-efficient performance, with sample complexity, runtime, and communication all bounded by roughly the square-root-scale term plus circuit depth. The authors note this extends earlier work that only handled label-invariant properties to general distribution properties. Source

4. Apple Compares Verifying Function Properties Against Distribution Properties

Apple. Apple published a theoretical study examining whether the known equivalence between testing location-invariant function properties and testing distribution properties carries over to the verification setting. Using doubly-sublinear interactive proofs of proximity, the researchers derived query bounds for uniform and frequency-bounded functions where both the verifier and prover read only a fraction of the input. They showed that the corresponding distribution properties have no doubly-efficient interactive proof, demonstrating that function and distribution properties diverge substantially once verification, rather than testing, is the goal. Source