VFlash predicts whole visual plan primitives in one pass instead of token by token
An NVIDIA paper replaces coordinate-by-coordinate decoding of points, boxes and trajectories with a drafter that emits each primitive at once, reporting fewer errors and faster generation.
The problem
Models for physical AI tasks, such as robotics and autonomous driving, can plan in more than language. The authors of VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding (arXiv 2611.08449) point to visual plans made of points, bounding boxes and trajectories, which represent spatial goals and anticipate motion.
The cost problem comes from how these plans are produced today. Autoregressive vision-language models generate them one coordinate token at a time, so richer predictions, such as long trajectories, become increasingly costly to decode.
The paper lists 19 authors and the organization NVIDIA. It was published on arXiv on 2026-11-05 and listed on Hugging Face daily papers on 2026-11-04.
What VFlash does
VFlash exploits the fact that visual primitives have a known structure. Instead of emitting coordinates one by one, it predicts the coordinates of a primitive jointly. A lightweight drafter generates an entire primitive in one forward pass.
The output layout is not fixed. It adapts to the primitive’s type, dimensionality and length, so the same mechanism covers compact boxes as well as long, variable-length trajectories.
The second idea concerns training and acceptance. The authors argue that spatial accuracy need not require reproducing an autoregressive model’s exact token sequence. VFlash therefore trains on annotated primitives and directly accepts structurally valid predictions without target-token agreement. In other words, a prediction is kept if it is a well-formed primitive, not because it matches what token-by-token decoding would have output.
Reported results
All figures below are the authors’ own, as given in the abstract:
- Across the evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduces prediction errors relative to autoregressive decoding.
- Preliminary single-rollout measurements indicate approximately 25x faster generation for 3D point motion forecasting.
- A world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline.
The authors themselves call the speed figure preliminary and based on single-rollout measurements.
What the abstract leaves open
The abstract does not say how large the error reductions are or which 2D and 3D trajectory tasks were evaluated. It does not name the vision-language model VFlash is attached to, the size of the drafter, or the hardware behind the speed measurement, and the roughly 25x figure covers only 3D point motion forecasting rather than boxes or 2D trajectories. For the MolmoSpaces results, it does not describe the baseline or the world action model in detail, or give per-suite numbers. It also does not discuss what happens when a structurally valid prediction is spatially wrong, since acceptance no longer checks agreement with the autoregressive model. No code repository is listed with the paper.
Source: VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding (arXiv 2611.08449)