Apple AI Updates: August 27, 2026

1. Apple Publishes PROOF-Gen for Tool-Calling Model Distillation

Apple. Apple researchers published PROOF-Gen, a distillation method that recovers failed teacher trajectories that standard generate-and-filter pipelines discard. A reflector analyzes execution traces and provides corrective guidance to steer teachers toward passing trajectories, then strips that guidance before student training. On the tau2-bench benchmark the method recovered 93% of failed scenarios and raised Qwen3-4B’s Pass@1 score from 0.132 to 0.529, with transfer to deployed on-device models across multiple locales. Source

2. Apple Introduces IDEA Prune Enlarge-and-Prune Pipeline

Apple. Apple researchers published IDEA Prune, an integrated pipeline that combines enlarged model pretraining with structured pruning under a unified learning rate schedule. The method uses an iterative pruning technique intended to reduce knowledge loss and redistribute model capacity. In experiments compressing 2.8B models to 1.3B, the approach outperformed training target-size models from scratch, indicating that enlarged pretraining helps even when the larger model is never deployed. Source

3. Apple Details Luce for Relightable 3D Asset Generation

Apple. Apple researchers published Luce, a 3D representation that generates relightable assets from a single image by unifying geometry and physically based rendering materials in a voxelized Gaussian cloud. It uses a variational autoencoder and a rectified-flow transformer to produce albedo, metallic-roughness, and normal maps. On the Toys4K benchmark, Luce improved FID by 28% over the strongest baseline while preserving fine details such as text and logos. Source