VSpark: Speculative Decoding for Autoregressive Vision-Language-Action Models
Abstract
VSpark accelerates autoregressive vision-language-action models by improving speculative decoding with calibrated action-noise rollouts, hidden-feature alignment, and reuse of rejected-draft continuations, achieving lossless 1.28x–2.10x speedups across multiple robot manipulation tasks.
Our aim is to accelerate autoregressive vision-language-action models (VLAs), which generate robot actions one token at a time. Speculative decoding makes this possible: a lightweight draft model proposes several tokens at once, and the VLA verifies them in a single pass, accepting only those that match its own predictions, so decoding remains lossless. The speedup, however, is limited in robot control, where many different tokens encode nearly the same motion and a draft close to the correct action is still rejected. To address this, our approach, VSpark, is built upon three design choices: (i) rollouts collected with calibrated action noise, which expose the draft to a wider range of robot states; (ii) feature alignment, which trains the draft to match the VLA's hidden features rather than its output tokens alone; and (iii) continuation reuse, which keeps the predictions the VLA has already computed along rejected drafts as candidates for the next verification step, instead of discarding them. On four VLAs with different action tokenizers and four manipulation benchmarks, VSpark achieves 1.28x to 2.10x speedup while keeping the VLA's decoding lossless.
Community
Our aim is to accelerate autoregressive vision-language-action models (VLAs), which generate robot actions one token at a time. Speculative decoding makes this possible: a lightweight draft model proposes several tokens at once, and the VLA verifies them in a single pass, accepting only those that match its own predictions, so decoding remains lossless. The speedup, however, is limited in robot control, where many different tokens encode nearly the same motion and a draft close to the correct action is still rejected. To address this, our approach, VSpark, is built upon three design choices: (i) rollouts collected with calibrated action noise, which expose the draft to a wider range of robot states; (ii) feature alignment, which trains the draft to match the VLA’s hidden features rather than its output tokens alone; and (iii) continuation reuse, which keeps the predictions the VLA has already computed along rejected drafts as candidates for the next verification step, instead of discarding them.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- VLN-Flash: Parallel Action Decoding with Adaptive Execution for Vision-and-Language Navigation (2026)
- POKE: Probe Offline, Keep Efficiently for Vision-Language-Action Models (2026)
- ReAT: Residual Action Tokenization for Autoregressive Robot Policies (2026)
- HAWK: Rethinking Multimodal Drafting for Speculative Decoding (2026)
- VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding (2026)
- APS-VLA: Action-Guided and Future-Aware Token Pruning for Efficient Vision-Language-Action Models (2026)
- When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Models citing this paper 1
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper