Apple AI Updates: July 17, 2026

1. Apple Shows Self-Distillation Improves Code Generation Without External Verifiers

Apple. Apple researchers described a self-distillation method that improves a language model’s code generation by sampling its own solutions at set temperature and truncation settings and then fine-tuning on those outputs, avoiding the need for external verifiers or teacher models. On Qwen3-30B-Instruct the approach raised pass@1 on LiveCodeBench v6 from 42.4 percent to 55.3 percent. The largest gains appeared on harder problems and held across multiple model sizes. Source

2. Apple Details Personalized Incremental Video Search for Apple TV

Apple. Apple published a personalization approach for Apple TV search that combines a text-based multilingual encoder trained with contrastive learning and an ID-based collaborative model derived from user interactions, feeding both similarity scores into an XGBoost ranker built from recent watch history. Offline testing showed improvements of 2.99 percent in NDCG@10 and 3.30 percent in MRR, with an 8.63 percent gain on ambiguous short queries. Online, the system produced statistically significant improvements of 1.14 percent in tap-through rate and 1.23 percent in conversion rate, with users who had longer viewing histories benefiting most. Source

3. Apple Introduces Doubly Sub-linear Interactive Proofs of Proximity

Apple. Apple researchers presented a proof system in which both the prover and the verifier read only a small portion of very large inputs, enabling verification of approximate claims about enormous datasets without accessing the full input. The authors construct such proofs for any property decidable by a constant-width read-once oblivious branching program, along with specialized proofs for Hamming weight approximation and graph bipartiteness variants. The work targets settings where reading an entire input is prohibitively expensive. Source