Apple AI Updates: October 13, 2026

1. SelfCompact Lets Agents Decide When to Summarize Their Own Context

Apple. Apple and Johns Hopkins researchers published SelfCompact, a NeurIPS paper in which an agent gets a compaction tool plus a short rubric describing when to compact (a sub-task is resolved) and when to hold off (mid-derivation or stuck). Across six competitive math and agentic search benchmarks and seven models, the method matched or beat fixed-interval summarization, improved on a no-summarization baseline by up to 18.1 points on math and 5 to 9 points on search, and cut per-question cost by 30% to 70%. It requires no fine-tuning, so it can be applied directly to existing agent scaffolds. Source

2. Study Finds LLM Translation Quality Annotation Remains Unreliable in Most Settings

Apple. An EMNLP paper from Apple compares LLM and human annotators under the MQM and ESA translation evaluation schemes on a long-context test set spanning 70 language pairs, plus WMT23 and WMT25 data. LLM agreement with humans exceeded human-to-human agreement on some tasks, but agreement varied widely by scheme, language pair, and domain, with LLMs struggling on minor errors, wrong language variants, and error span marking. The authors recommend human-LLM collaborative annotation pipelines for teams using LLMs as translation judges. Source

3. Minions Pools Idle Apple Silicon Macs Into a Distributed Inference Cloud

Apple. Apple researchers described Minions, a system that turns idle Apple Silicon GPUs across macOS devices into a virtual GPU cloud for large-scale ML inference, with flexible task execution and fault tolerance. The work is aimed at teams facing cost, elasticity, and queueing limits on centralized GPU infrastructure, and the paper is available as arXiv 2610.12970. Apple’s summary does not report quantitative throughput results. Source