VSpark: Speculative Decoding for Autoregressive Vision-Language-Action Models
Paper • 2611.22666 • Published • 5
Draft models for VSpark: Speculative Decoding for Autoregressive Vision-Language-Action Models. VSpark keeps verification exact for lossless decoding.
Paper: arXiv:2611.22666
Each draft must be paired with its target:
| File | Target | LIBERO speedup |
|---|---|---|
vspark_pi0_fast.pt |
lerobot/pi0fast-libero | 1.93× |
vspark_actioncodec.pt |
ZibinDong/SmolVLM2-2.2B-ActionCodec-AR-LIBERO | 2.10× |
vspark_starvla.pt |
StarVLA/Qwen2.5-VL-FAST-LIBERO-4in1 | 1.61× |
Speedups are averaged over the four LIBERO suites on one RTX 4090.
Each draft contains the action-token rows of its target's output head, so it follows the license of that target:
| File | License |
|---|---|
vspark_pi0_fast.pt |
Gemma Terms of Use (PaliGemma) |
vspark_actioncodec.pt |
Apache License 2.0 (SmolVLM2) |
vspark_starvla.pt |
Qwen Research License (Qwen2.5-VL-3B), non-commercial |
The VSpark code is released under the Apache License 2.0.
huggingface-cli download suhoyoo/VSpark --local-dir checkpoints
python inference/eval_libero.py --model pi0_fast --suite libero_goal \
--checkpoint checkpoints/vspark_pi0_fast.pt --output outputs/libero_goal
See the code repository for installation and training.
@article{yoo2026vspark,
title = {VSpark: Speculative Decoding for Autoregressive Vision-Language-Action Models},
author = {Yoo, Suho and Jang, Youngjoon and Cho, Hyebin and Lee, Taeyeop and Jeong, Jaewook and Chung, Joon Son},
journal = {arXiv preprint arXiv:2611.22666},
year = {2026}
}