Apple AI Updates: May 9, 2026

1. Apple proposes RVPO for risk-sensitive RLHF alignment

Apple. RVPO is a multi-objective RLHF aggregation method that penalizes inter-reward variance during advantage computation, pushing the policy toward consistent performance across all reward heads instead of letting it trade safety for helpfulness on the average. The framing matters for production alignment stacks where one objective (safety, refusal calibration) is regularly drowned out when the helpfulness reward dominates. Source

2. HeadsUp scales 3D Gaussian head reconstruction from multi-view captures

Apple. HeadsUp uses an encoder-decoder over UV-parameterized Gaussians so reconstruction quality scales with input resolution without scaling the number of 3D Gaussians. The architecture targets large-scale, high-fidelity head capture pipelines where prior Gaussian-splatting heads either blew up memory or lost detail at high input resolution. Source

3. Velox learns 4D representations for dynamic geometry and appearance

Apple. Velox introduces “dynamic shape tokens” — a compact latent representation of moving 3D objects — trained jointly with separate decoders for geometry and appearance so the same latent serves rendering and reconstruction. The pitch is a representation that is descriptive enough to drive downstream tasks while compressive enough to ship to device. Source

4. Apple announces 2026 Privacy-Preserving ML & AI Workshop

Apple. Apple opened its 2026 workshop on privacy-preserving machine learning, framing privacy as a research constraint baked into model design rather than a post-hoc deployment concern. The event is the venue Apple uses to surface federated learning, differential privacy, and on-device inference work, and is worth tracking for teams shipping models against PPML constraints. Source