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.
Figures & tables
Figure 2: Overview of VSpark. During training, (a) rollouts with calibrated action noise collect target supervision at perturbed states, and (b) feature alignment conditions and supervises the draft with target representations. During inference, (c) continuation reuse guides the next draft tree with target continuations along rejected branches. Strict verification checks all candidates under their current prefixes, preserving lossless decoding.
Figure 3: Illustration of continuation reuse. During verification in cycle t , the target also predicts along rejected branches. VSpark reuses them to build the tree for cycle t+1 , where they are re-verified.
Goal
Spatial
Object
Long
Avg.
Method
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
π0 -FAST
EAGLE3
3.09
1.22×
3.12
1.24×
3.31
1.31×
3.17
1.27×
3.17
1.26×
P-EAGLE
2.94
1.45×
2.97
1.49×
3.12
1.57×
3.00
1.48×
3.01
1.50×
DFlash
2.33
1.39×
2.42
1.46×
2.44
1.46×
2.30
1.36×
2.37
1.41×
DFlash2
2.99
1.50×
3.14
1.56×
3.20
1.59×
2.96
1.47×
3.07
1.53×
DSpark
3.08
1.57×
3.12
1.59×
3.31
1.65×
3.14
1.60×
3.16
1.60×
Table 1: Main results on four LIBERO suites. VSpark achieves the highest acceptance length τ and speedup on all four VLAs and all suites.
Spoon
Carrot
Stack
Eggplant
Avg.
Method
τ↑
Spd. ↑
τ↑
Spd. ↑
τ↑
Spd. ↑
τ↑
Spd. ↑
τ↑
Spd. ↑
EAGLE3
3.89
1.69×
3.49
1.59×
3.68
1.64×
3.84
1.66×
3.72
1.65×
P-EAGLE
3.37
1.70×
2.95
1.53×
3.08
1.57×
3.26
1.62×
3.16
1.61×
DFlash
2.85
1.51×
2.52
1.40×
2.59
1.41×
2.62
1.38×
2.64
1.42×
DFlash2
3.21
1.62×
2.89
1.51×
2.97
1.51×
3.00
1.50×
3.02
1.54×
DSpark
3.70
1.83×
3.31
1.72×
3.39
1.71×
3.50
1.74×
3.48
1.75×
Table 2: SimplerEnv results with StarVLA-FAST. Tasks: spoon on towel, carrot on plate, stack cubes, eggplant in basket. Spd.: speedup.
LIBERO-Plus
LIBERO-Pro
Goal
Spatial
Object
Long
Goal
Spatial
Object
Long
Avg.
Method
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
π0 -FAST
EAGLE3
2.74
1.00×
2.78
1.01×
2.94
1.07×
2.78
1.00×
2.95
1.45×
3.03
1.48×
3.29
1.58×
3.06
1.46×
2.95
1.26×
P-EAGLE
2.60
1.40×
2.63
1.39×
2.77
1.49×
2.64
1.33×
2.79
1.69×
2.88
1.76×
3.09
1.81×
2.89
1.75×
2.79
1.58×
DFlash
1.92
1.23×
1.94
1.25×
1.92
1.25×
1.89
1.18×
2.25
1.68×
2.36
1.72×
2.39
1.73×
2.26
1.66×
2.12
1.46×
DFlash2
2.46
1.32×
2.53
1.37×
2.62
1.42×
2.44
1.29×
2.73
1.68×
2.96
1.78×
3.07
1.80×
2.80
1.67×
2.70
1.54×
Table 3: Main results on LIBERO-Plus and LIBERO-Pro. Both benchmarks evaluate VLAs under diverse variations in visual appearance, object configurations, robot states, language instructions, and task conditions. VSpark consistently achieves the highest acceptance length and inference speedup across all eight suites and four VLAs. Avg. denotes the arithmetic mean across the eight suites.
Distance
Relaxed ratio
τ↑
Speedup ↑
SR (%) ↑
-
0.00%
3.16
1.60×
91.7
0.011
2.05%
3.18
1.62×
91.7
0.012
12.70%
3.31
1.67×
91.0
0.013
23.72%
3.47
1.74×
89.2
0.015
35.79%
3.70
1.82×
83.8
0.020
45.97%
4.02
1.90×
71.4
Table 4: Embedding-based relaxed verification on LIBERO with π0 -FAST and DSpark drafts, averaged over four suites. A draft token is accepted when its cosine distance from the target token embedding is at most the given distance. Relaxed ratio is the fraction of accepted draft tokens that differ from the target prediction.
Figure 4: Token embedding distance versus action discrepancy for π0 -FAST (left) and ActionCodec (right). Lines show the median and shaded regions the range of the 6D action RMSE at each distance. Similar distances span a wide range of action errors.
π0 -FAST
ActionCodec
Reject Pos.
Draft
Target
Δ
Draft
Target
Δ
1
24.70
39.50
+14.80
35.55
54.24
+18.69
2
35.20
45.96
+10.76
38.54
60.05
+21.51
3
28.20
55.12
+26.92
45.14
65.48
+20.35
4
29.26
53.40
+24.14
32.68
47.75
+15.07
5–7
46.57
56.42
+9.85
36.22
54.35
+18.13
Table 5: Next-token agreement after rejection (%), over four LIBERO suites. Draft and Target: how often the next token each predicts after the rejected token matches the target’s next token after the correction. Δ : their difference in percentage points.
π0 -FAST
ActionCodec
1 Layer
5 Layers
1 Layer
5 Layers
Chain
Tree
Tree
Chain
Tree
Tree
Method
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
EAGLE3
2.28
1.04×
3.17
1.26×
3.35
0.70×
2.42
1.13×
3.99
1.63×
4.11
0.98×
P-EAGLE
1.99
1.16×
3.01
1.50×
3.11
1.17×
2.36
1.36×
3.83
1.95×
4.04
1.67×
DFlash
1.55
1.05×
2.37
1.41×
2.73
1.20×
2.17
1.37×
3.36
1.90×
3.27
1.51×
Table 6: Effect of verification strategy and draft depth. Tree verification consistently improves acceptance length and speedup over chain verification. Increasing draft depth to five layers generally improves acceptance length but reduces speedup.
Figure 5: Conditional token acceptance in VSpark for π0 -FAST (left) and ActionCodec (right): acceptance rate of candidates promoted by continuation reuse (Fusion) and of the others, among candidates whose entire parent path has been accepted.
π0 -FAST
ActionCodec
Method
τ↑
Speedup ↑
τ↑
Speedup ↑
DSpark
3.16
1.60×
3.69
1.93×
+ Continuation Reuse
3.22
1.69×
4.04
2.01×
+ Feature Alignment
3.87
1.90×
4.10
2.07×
+ Action Noise (VSpark)
3.94
1.93×
4.21
2.10×
Table 7: Component ablation of VSpark. Each component progressively improves acceptance length and inference speedup on both VLAs. Feature alignment provides the largest improvement on π0 -FAST, while continuation reuse contributes most on ActionCodec. Combining all three components achieves speedups of 1.93× and 2.10× , respectively.
Table A.1 : Evaluation episodes per setting. Counts are reported for one target model, method, and setting before error exclusions. Each benchmark follows its official evaluation protocol.
Target
Baselines (no noise)
VSpark (noise)
π0 -FAST
27,548
27,548
StarVLA-FAST
27,548
27,548
ActionCodec
27,548
27,548
G0.5
27,548
27,548
Appendix
Table A.2 : Draft training set sizes. Baselines use rollouts without action noise; VSpark uses rollouts with calibrated action noise. Sample counts and optimizer updates are matched.
Model / method
Repository / checkpoint
Target models
π0 -FAST [ 40 ]
lerobot/pi0fast-libero
StarVLA-FAST (LIBERO) [ 12 ]
StarVLA/Qwen2.5-VL-FAST-LIBERO-4in1
checkpoints/steps_30000_pytorch_model.pt
StarVLA-FAST (WidowX) [ 12 ]
StarVLA/Qwen-FAST-Bridge-RT-1
checkpoints/steps_10000_pytorch_model.pt
Appendix
Table A.3 : Model checkpoints and baseline repositories. Identifiers link to Hugging Face for target models and GitHub for baselines. LIBERO checkpoints are shared across LIBERO, LIBERO-Plus, and LIBERO-Pro; the WidowX checkpoint is used for SimplerEnv.
Target
π0 -FAST
StarVLA-FAST
ActionCodec
G0.5
Target size
3B
3B
2.2B
3B
VSpark Draft size
205M
147M
119M
150M
Layers
1
1
1
1
Hidden dim.
2,048
2,048
2,048
2,048
Query heads
8
16
32
8
KV heads
1
2
32
2
Appendix
Table A.4 : Draft specifications. Target sizes follow the public model labels; draft sizes are rounded to the nearest million parameters. KV denotes key and value heads; FFN dim. is the intermediate dim.
Target
Feature layers
π0 -FAST
2, 5, 9, 13, 16
StarVLA-FAST
2, 10, 18, 26, 34
ActionCodec
2, 7, 12, 17, 22
G0.5
2, 7, 12, 17, 22
Appendix
Table A.5 : Target layers for feature conditioning. Indices refer to Transformer block outputs and start at one. Each target supplies five layers spanning shallow to deep representations.
Translation (mm)
Rotation ( ∘ )
Target
σx
σy
σz
σrx
σry
σrz
π0 -FAST
3.88
7.10
3.08
0.24
0.66
0.90
StarVLA-FAST
4.88
6.49
6.17
0.42
1.36
1.67
ActionCodec
2.94
2.73
3.31
0.33
0.49
0.34
G0.5
1.17
0.87
2.30
0.22
0.39
0.38
Appendix
Table A.6 : Calibrated action noise on LIBERO. We report the marginal standard deviations of the fitted zero-mean Gaussian, converted to action-command perturbations in millimeters and degrees. Each model uses 64 calibration inputs, with 16 from each suite.
Feature input
Regression
τ↑
Speedup ↑
×
×
3.23
1.68×
✓
×
3.27
1.70×
×
✓
3.67
1.78×
✓
✓
3.94
1.93×
Appendix
Table B.1 : Feature alignment component ablation on π0 -FAST. We independently enable feature conditioning and feature regression. All settings use one-layer drafts, training only on noise rollouts, continuation reuse, and early stopping. Results are averaged over LIBERO suites.
Reuse
Early stop
τ↑
Speedup ↑
×
×
3.87
1.84×
✓
×
3.94
1.86×
×
✓
3.87
1.86×
✓
✓
3.94
1.93×
Appendix
Table B.2 : Continuation reuse component ablation on π0 -FAST. We independently enable continuation reuse and early stopping. All settings use one-layer drafts, training only on noise rollouts, and feature alignment. Results are averaged over LIBERO suites.
Method
≥1
≥2
≥3
≥4
≥5
≥6
≥7
τ↑
π0 -FAST
DSpark
91.63
79.42
32.00
10.43
1.87
0.54
0.25
3.16
VSpark
95.19
81.26
54.76
33.94
16.71
7.91
3.96
3.94
ActionCodec
DSpark
83.34
66.89
47.30
31.25
20.82
12.32
6.97
3.69
VSpark
92.80
77.98
58.29
39.83
26.83
15.93
9.31
4.21
Appendix
Table B.3 : Statistics of accepted draft length. We report the percentage of verification cycles accepting at least k consecutive draft tokens. Mean committed length τ additionally includes target correction or additional tokens. Results are averaged over LIBERO suites.
π0 -FAST
ActionCodec
Replacement probability
Online
Post-hoc
Online
Post-hoc
0% (AR)
90.6
90.6
85.4
85.4
1%
83.2
79.4
86.4
87.0
5%
59.4
52.0
79.2
81.0
10%
35.0
30.6
68.0
69.8
25%
7.6
9.4
29.0
29.2
Appendix
Table B.4 : Nearest-embedding token replacement without drafting. Each eligible action token is replaced with probability p by its nearest distinct token in the action vocabulary under cosine distance. Online replacement conditions subsequent generation on the modified prefix; post-hoc replacement modifies the completed AR chunk before action decoding and execution. We report success rates (%) on LIBERO Goal.
Goal
Spatial
Object
Long
Avg.
Method
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
τ↑
Speedup ↑
π0 -FAST
EAGLE3
3.09
1.22×
3.12
1.24×
3.31
1.31×
3.17
1.27×
3.17
1.26×
+ VSpark
4.31
1.60×
4.48
1.65×
4.59
1.71×
4.34
1.57×
4.43
1.63×
P-EAGLE
2.94
1.45×
2.97
1.49×
3.12
1.57×
3.00
1.48×
3.01
1.50×
+ VSpark
3.65
1.74×
3.78
1.78×
3.82
1.79×
3.64
1.70×
3.72
1.76×
Appendix
Table B.5 : Applying VSpark to other draft architectures on LIBERO. We train each baseline on rollouts with calibrated action noise and add feature alignment, continuation reuse, and early stopping. Speedup uses the same target autoregressive reference for each baseline and its VSpark variant. Avg. is the arithmetic mean over LIBERO suites.
π0 -FAST
ActionCodec
Component
Scope
Ratio
P(eK>eD)
Ratio
P(eK>eD)
Translation
First step
2.68 ×
85.00%
17.76 ×
97.04%
Full chunk
4.56 ×
94.69%
15.28 ×
99.61%
Rotation
First step
3.03 ×
87.44%
18.98 ×
96.38%
Full chunk
4.66 ×
96.30%
10.11 ×
99.45%
Appendix
Table B.6 : Kalman prediction errors versus draft action discrepancies. We compare Kalman prediction errors with action differences induced by replacing a single rejected token while keeping the remaining target tokens fixed. Results are reported separately for translation and rotation, at the first timestep and across the full action chunk. Ratio denotes the median of eK/eD , where eK is the Kalman prediction error and eD is the action discrepancy caused by token replacement. P(eK>eD) denotes the fraction of cases where the Kalman error exceeds the draft discrepancy.
BF16
FP32
Suite
AR SR [95% CI]
VSpark SR [95% CI]
Agreement
AR SR
VSpark SR
Agreement
π0 -FAST
Goal
91.64 [89.48, 93.80]
91.40 [90.16, 92.64]
88.56
91.00
91.00
100.00
Spatial
94.92 [93.89, 95.95]
95.48 [94.37, 96.59]
94.80
95.80
95.80
100.00
Object
97.40 [97.15, 97.65]
96.68 [96.35, 97.01]
95.68
96.40
96.40
100.00
Long
86.87 [86.11, 87.63]
85.33 [82.34, 88.33]
82.47
85.20
85.20
100.00
Appendix
Table B.7 : Numerical precision and agreement with autoregressive decoding on LIBERO. All values are percentages. BF16 reports mean SR with 95% Student- t confidence intervals. Agreement is the fraction of paired episodes with the same success/failure outcome under AR and VSpark. Avg. averages the four suites within each seed before computing the mean and confidence interval.
Method
LIBERO
LIBERO-Plus
LIBERO-Pro
Target AR
0.35%
1.65%
1.14%
EAGLE3
0.33%
1.62%
1.18%
P-EAGLE
0.45%
1.63%
1.10%
DFlash
0.38%
1.63%
1.04%
DFlash2
0.48%
1.64%
1.16%
DSpark
0.38%
1.55%
1.20%
Appendix
Table B.8 : Generation-error rates across evaluation methods. For each method, we report the fraction of episodes containing generation errors in its own rollouts, including failures to decode generated tokens into valid actions. Values are averaged over target models and benchmark suites. Task failures and token-limit hits alone are not counted as generation errors.
Method
Draft
Target
Prefill
Action
Other
Total
AR
-
12.02
2.40
0.02
0.03
14.47
DSpark
1.58
4.23
2.41
0.02
0.70
8.93
VSpark
1.32
3.30
2.24
0.02
0.56
7.44
Appendix
Table B.9 : Time breakdown of π0 -FAST on LIBERO. Times are ms per output token, averaged over four suites. Draft includes tree construction; Target denotes AR decoding or speculative verification; Action denotes action decoding.
Configuration
τ↑
Speedup ↑
VSpark
3.94
1.93×
w/o Action Noise
3.87
1.90×
w/o Feature Alignment
3.23
1.68×
w/o Continuation Reuse
3.87
1.84×
DSpark
3.16
1.60×
Appendix
Table B.10 : Independent component removal on π0 -FAST. Each variant removes one component from VSpark. Continuation reuse includes early stopping. Results are averaged over four LIBERO suites.
Benchmark
Method
τ↑
Speedup ↑
Acc. ↑
Qwen3-4B
GSM8K
AR
—
1.00×
91.89
DSpark
7.455
4.44×
91.51
+ Reuse
7.454
4.44×
91.74
HumanEval
AR
—
1.00×
83.54
DSpark
7.089
4.28×
85.37
Appendix
Table B.11 : Continuation reuse on language models. We report acceptance length, speedup relative to AR, and accuracy (%). + Reuse adds continuation reuse to DSpark under the same tree budget.
Method
Verification
τ↑
Speedup ↑
SR
AR
-
1.00
1.00×
76.2
SpecVLA
Strict
1.70
0.99×
77.6
SpecVLA
Relaxed
2.38
1.23×
75.0
KERV
Relaxed
2.36
0.96×
73.2
DSpark
Strict
1.80
1.10×
74.2
DSpark
Relaxed
3.19
1.59×
76.8
Appendix
Table B.12 : OpenVLA on LIBERO Goal. Strict uses exact verification; Relaxed uses the acceptance rules described in the text. SR is in percent. A dash denotes an unreported result.
Vision-language-action (VLA) policies generate robot controls autoregressively, making closed-loop latency dominated by repeated target-model forward passes. Speculative decoding reduces this cost by verifying blocks of draft action tokens in parallel, and recent VLA methods further relax token-level acceptance because small differences in action-token space often map to similar continuous controls. However, this relaxation remains scene-agnostic. A fixed token-distance tolerance treats the same action-token deviation as equally safe across states, although deviations that are harmless in free space can cause collisions or grasp failures near contact. We propose WA-SpecDec, a world-aware speculative decoding framework that injects world-model-derived physical scene awareness during the VLA prefill stage, producing shared world-aware prefill states for draft proposal and target verification without changing the relaxed acceptance rule. Across three state-of-the-art relaxed acceptance schemes, WA-SpecDec preserves higher task success under looser relaxation and enables longer accepted prefixes. At comparable-success operating points, WA-SpecDec achieves a 1.5x matched-success speedup over VLA speculative decoding alone and reduces near-contact failure (NCF) by 18.6% on average relative to the corresponding speculative baselines.
Test-time sampling can improve Vision-Language-Action (VLA) policies, but only if the model can efficiently select a good action from multiple candidates. Prior work often relies on separately trained external verifiers, adding extra models and inference overhead. We propose Token Probability Bucketing (TokenPB), a self-verification approach that discretizes continuous action chunks into bucket tokens and co-trains a single VLA checkpoint with a joint objective: a flow-matching loss for action generation and an auxiliary bucket loss Lbucket that teaches the VLM backbone to model observation-conditioned distributions over bucket-token sequences. At inference, TokenPB samples M candidates, scores each via teacher-forced evaluation under the learned distribution, and selects the best---enabling efficient batched scoring with observation-prefix KV-cache reuse. We provide theory connecting the training objective to a monotonic ranking property and bounds on expected selection regret that separate candidate-set effects from calibration. Across multiple backbones and benchmarks, TokenPB yields consistent success-rate improvements over single-sample inference from the same checkpoint (+2.7 in simulation, +17.1 in real-world, +16.4 under out-of-distribution shifts), and further improves over external-verifier baselines under the same sampling budget. The code is publicly available at https://github.com/RUCKBReasoning/TokenPB.
Chen Zhao, Zhuoran Wang, Haoyang Li +6
School of Information, Renmin University of China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China · Beijing University of Posts and Telecommunications +1
Vision-Language-Action (VLA) models exhibit strong generalization for robotic manipulation, yet their high inference latency limits real time deployment. We identify two primary sources of temporal redundancy in existing VLA pipelines: repeated visual encoding of highly similar consecutive frames and multi step iterative sampling in diffusion based policies. To address this, we propose a system level acceleration strategy that reduces computation in both perception and action generation. On the perception side, we incrementally update only tokens corresponding to dynamic scene regions instead of re-encoding entire frames. On the policy side, we compress diffusion sampling into a compact 2-step schedule through efficiency oriented training while preserving action precision. Experiments on Libero, RobotWin, and Real Robot Platforms demonstrate over 2 times speedup while maintaining high performance, achieving up to 98% success rate on general manipulation benchmarks. Our codes will be released on Github.
Yuzhou Wu, Yuxin Zheng, Muchun Niu +6
1Tianji KernalMind co ltd · 2Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · 3Shanghai Jiao Tong University, Shanghai, China +2