Apple AI Updates: August 29, 2026

1. Apple Studies Whether LLMs Update Beliefs Like Bayesian Reasoners

Apple. Apple researchers published work measuring how far large language models deviate from optimal Bayesian belief updates when incorporating new evidence, introducing an “information processing gap” metric to quantify the inconsistencies. The study found that non-Bayesian heuristic updates often outperform exact Bayesian updates on downstream tasks, indicating that the models operate with misspecified internal world representations. The team evaluated multiple evidence-incorporation approaches and showed how the measurements can diagnose failures in LLM-powered inference systems. Source

2. Apple Introduces Agent Seer for Synthesizing Agent Evaluation Scenarios

Apple. Apple researchers published Agent Seer, a pipeline that automatically generates realistic evaluation scenarios for AI agents using only tool specifications, without manual curation, example data, or live tool access. The system enriches raw schemas, generates graded scenarios with synthetic tool outputs, and expands them into mock-data-grounded multi-turn dialogues. Testing across seven Model Context Protocol specifications found that parameter schema complexity most strongly correlates with quality variation and that argument value accuracy is the primary failure mode, while the pipeline achieved complete tool coverage on small and medium specifications. Source