Core ML, Applied Evaluation, and Efficient Systems
Efficiency leads include LoopFill’s reported 5.85×/5.13× speculative-decoding speedups and Latent Feedback’s reported compute/data savings, alongside Loopback Transformer and MMFlash results; scale, workload, and replication remain open. Accelerate Bench shows application-level optimization scores vary widely and identical code can run 0.29×–2.12× depending on build/runtime context. UMM-JEPA reports a visual-understanding gain over its baseline, while transfer and independent validation remain important.
Sources (36)
Updated Nov 18, 2026