Feng Han

Robotic world-action models do not need pretrained internet video backbones to scale: training native planning and dynamics modules from scratch across 30,000 hours unlocks strong zero-shot manipulation and drives a 17.7-point gain on an embodiment comprising under 2% of the data.

Monolithic video evaluation dilutes critical action cues, but chunking interactive rollouts into action-aligned visual evidence allows targeted reward models to outperform GPT-5.5 at scoring world model dynamics.

Stale content complaints dropped by nearly 55% in Google Discover after implementing a dual-filter framework that intelligently targets content relevance decay and supersession.

Achieving lossless processing of 256K contexts, Keye-VL-2.0 transforms how we approach long-video understanding and agentic intelligence.