Fine-grained Computation-Communication Overlap via Tile-level Signaling and Scheduling for Mixture-of-Experts

This RSS batch contains multiple systems and ML-systems papers with operational and privacy-relevant implications. Key items: SynPre-FL (synthetic-data pretraining for federated learning) claims preservation against membership-inference and reconstruction attacks but raises the usual caveats about threat modeling and potential residual disclosure risk from synthetic generators or downstream model leakage; “Do Co-Located AI Training Jobs Synchronize?” identifies an emergent availability risk where shared power caps and load-dependent throttling can phase-lock training jobs, producing large, amy