VFlash finding: lightweight drafter can beat its autoregressive target on decoding accuracy
ZhidingYu · x · 2026-11-06
Counterintuitive finding from the VFlash thread: a lightweight drafter can outperform its autoregressive target on decoding accuracy — contradicting the LLM speculative decoding assumption that AR is the accuracy upper bound, so verification can be skipped in some tasks. The author speculates visual intelligence may be fundamentally parallel prediction while LLM intelligence is autoregressive. Physical AI decoding is characterized by rich output types (text planning plus visual primitives), structured outputs (points, boxes, trajectories), and versatile decoding requirements.
More from Embodied
- Travis Kalanick's new firm Atoms builds AI and robots for food, mining, transport — davidsenra · 2026-11-06
- Verne V0.5 trains a robot policy with video prediction, but runs actions-only — rohanpaul_ai · 2026-11-06
- Robot success rate doubles 23%→51% from pretraining data alone, no fine-tuning — rohanpaul_ai · 2026-11-06
- Rabbit R1 with OS3 actually handles a hospital equipment return — jesselyu · 2026-11-06
- Do Heavy Robot Arms Hurt Policy Performance? New Thread Weighs the Hidden Costs — pulkitology · 2026-11-06
- Dream-1 updated with single-arm SO-101 and single-arm Franka support — scott_e_reed · 2026-11-06