Apple AI Updates: October 1, 2026

1. Apple’s SCLATE Compresses Month-Long Agent Scenarios Into Hours for Continual-Learning Research

Apple. SCLATE is an execution substrate for training and evaluating agents across many sessions, using an event scheduler and a hybrid simulated clock that runs in real time while the agent works and skips idle gaps. Apple ported seven benchmarks and tested ten agent configurations across ten models, finding that add-on memory systems did not consistently beat the harness’s native memory. Post-training Qwen3.5-4B on SCLATE rollouts cut file line reads by 6.8x and raised its SWE-bench Verified pass rate by 16.7 points, which is relevant to teams deciding whether to bolt external memory onto coding agents. Source

2. Apple Found Activation Steering Works Far Worse on Instruction-Tuned Models

Apple. In an EMNLP paper, Apple researchers compared activation steering, prompting, and supervised fine-tuning for injecting and removing concepts from LLM outputs, scoring both effectiveness and fluency. Cheap steering methods often achieved conditioning only at a steep cost to fluency, and activation steering was far less effective on instruction-tuned models than on their base versions. The authors also found that inexpensive textual metrics correlated strongly with costly LLM-as-judge evaluations, a useful shortcut for anyone benchmarking steering methods. Source