Apple AI Updates: July 21, 2026

1. Apple Proposes RayRoPE for Multi-View Attention

Apple. Apple researchers introduced RayRoPE, a projective ray positional encoding scheme for multi-view transformers. It encodes image patches uniquely and enables SE(3)-invariant attention across multiple camera views, targeting 3D and multi-view vision tasks where consistent geometry across viewpoints matters. Source

2. Apple Releases LVSum, a Timestamp-Aware Long Video Summarization Benchmark

Apple. Apple published LVSum, a benchmark for evaluating long-form video summarization with fine-grained temporal alignment. It spans 72 videos and emphasizes timestamp-aware summarization quality, giving researchers a way to measure how well models locate and describe events across long videos rather than producing loosely grounded summaries. Source

3. Apple’s Length Value Model Adds Token-Level Length Awareness

Apple. Apple proposed the Length Value Model, a framework that models the remaining generation length at each decoding step to give language models token-level awareness of how much output is left. The approach relies on scalable value pretraining and aims to improve control over output length, a persistent weakness in instruction-following and structured generation. Source