Apple AI Updates: September 18, 2026

1. REVERSAL-BENCH Measures the Point Where Reset-Free RL Falls Apart

Apple published a benchmark that isolates reversibility as an axis in autonomous reinforcement learning. Reset-free RL aims to train policies without external resets, but the paper argues existing setups quietly assume the environment is reversible, a property real-world manipulation lacks once an agent pushes an object off a table or spills a container. REVERSAL-BENCH pairs a reversibility axis with a reset oracle so the drop in performance as irreversibility increases can be measured rather than assumed away. Source

2. Apple Reframed Flow-Matching Distillation as an Energy Navigation Problem

Apple researchers targeted the cost of discrete flow matching, where generating text can take hundreds of forward passes. Standard distillation trains a student to reproduce the teacher’s multi-step trajectory in a few steps, and when the student underperforms the usual explanation blames student capacity. The paper treats the trajectory itself as the teaching signal and uses energy-navigated distillation to guide the few-step student instead. Source