Apple AI Updates: July 22, 2026

1. Apple Generates Agent Training Data Without Executable Environments

Apple. Apple researchers proposed an environment-free method for producing synthetic training trajectories for API-calling agents, avoiding the need to build fully implemented environments with executable APIs and populated databases. Given only API specifications, the system uses an LLM as a digital world model to generate diverse tasks, a teacher agent to solve them while an LLM simulates API responses, and an LLM judge to filter for quality. Models fine-tuned on the resulting data showed significant gains on the AppWorld and OfficeBench benchmarks, pointing to a more scalable path for training tool-using agents. Source

2. Apple’s CalibAtt Speeds Up Text-to-Video Diffusion With Sparse Attention

Apple. Apple introduced CalibAtt, a training-free technique that accelerates text-to-video diffusion models by skipping low-value attention connections. An offline calibration phase identifies token-to-token connections that consistently yield negligible attention scores, then compiles optimized sparse-attention operations for each layer, head, and diffusion timestep so only meaningful connections are computed at inference. On models including Wan 2.1 14B and Mochi 1, the method reached up to 1.58x end-to-end speedup while preserving video quality and text-video alignment. Source