Verify, Act, Distill: Turning Test-Time Compute into Test-Time Adaptation for Vision-Language-Action Models

Kaneyoshi Hiratsuka; Satoshi Yamamori; Jun Morimoto

cs.ROOct 10, 2026

Verify, Act, Distill: Turning Test-Time Compute into Test-Time Adaptation for Vision-Language-Action Models

Authors: Kaneyoshi Hiratsuka, Satoshi Yamamori, Jun Morimoto

Organizations: Kyoto University

Abstract

Test-time compute (TTC), which scores multiple candidate actions at inference time, is a powerful way to improve the execution performance of vision-language-action (VLA) models, but existing methods keep the policy frozen, so the gain must be recomputed at every execution. We propose Verify, Act, Distill, a framework that consolidates the gains of TTC into the VLA itself. A fixed verifier scores candidates generated by a flow-matching VLA, the weighted aggregate is executed, and the observations and aggregated chunks from successful adaptation episodes are distilled directly into LoRA parameters with the standard flow-matching loss. This reuses supervision that TTC has already produced during execution, without additional reinforcement learning or rollouts dedicated to teacher generation. On the 40 LIBERO tasks with a SmolVLA 5k checkpoint, distilling from only 10 episodes per task raises the mean success rate of verifier-free single-candidate generation from 54.1% to 70.6%, retaining about 74% of the improvement obtained by aggregating eight candidates. Applying the same aggregation to the distilled policy further raises the success rate from 76.4% to 82.3%, with a particularly large gain on LIBERO-Long (44.0% to 57.8%), and the single-candidate policy also improves on LIBERO-plus, which contains distribution shifts. These results show that using verification as a learning signal, rather than only as an execution-time selector, both retains the improvements of TTC and re-amplifies the effect of subsequent TTC.

Figures & tables

Figure 1: Overview of Verify, Act, Distill. Candidate chunks are scored and aggregated by the verifier and the leading $H_{2}=10$ steps are executed. Observations and the aggregated chunks of length $H_{1}=50$ from successful episodes are stored in a buffer and used to update the LoRA parameters. The verifier evaluates only the first $H_{2}$ steps of each candidate, whereas the teacher data include the entire planned chunk.

First 35 tasks per suite (common subset)
Method (10k, LIBERO-plus)	Spatial	Object	Goal	Long	Avg.
Base	35.5	18.8	13.0	16.2	20.9
Agg-8 ( $V_{\mathrm{act}}$ )	41.5	24.9	14.3	21.7	25.6
Base	36.6	15.1	14.9	16.0	20.7
Agg-8 ( $V_{\mathrm{act}}$ )	40.6	17.7	17.1	23.4	24.7
Distill	41.1	17.7	16.6	22.6	24.5

Table 2: Success rates (%) on LIBERO-plus, SmolVLA 10k, 105 tasks per suite, seeds 10–19 (

n{=}1050

per suite). The lower block uses the first 35 tasks per suite (140 tasks,

n{=}350

), on which Base and Agg-8 are re-aggregated for comparison.

Panel of Figure 3 — Figure 3: Dependence on the number of candidates and the adaptation budget. Both use the variant without the success filter; numbers are in Tables 6 and 7 .

Figure 4: Rollouts on LIBERO-Long task 9 (SmolVLA 5k) from the same initial state (seed 12) for Base ( $N{=}1$ ), Agg-8, and distilled policy + Agg-8; thin lines show each candidate’s predicted trajectory at each chunk boundary and the thick line the executed weighted average, and rows show the progression of time. The seed is the first one, searching upward from seed 10, in which Base fails and the distilled policy + Agg-8 succeeds (Agg-8 alone also fails here).

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Trained component	Data and supervision
SmolVLA (BC, 20k steps)	Only the 5,225 successful episodes among the newly collected rollouts. Official LIBERO demonstrations not used.
Verifier	Eight candidates generated at states from both the official LIBERO demonstrations (all four suites) and the newly collected successful rollouts. Supervision is the RMSE to the $H_{2}=10$ recorded action steps at each state. The 10k candidate dataset comprises 104,294 states.
World model (appendix)	Official LIBERO demonstrations and newly collected rollouts. For rollouts, both successful and failed data are used.

Appendix

Table 3: Breakdown of offline training data by use. This is distinct from the success-filtered distillation buffer used in the adaptation phase.

Teacher	Filter	Verifier	Spatial	Object	Goal	Long	Avg.
Mean	On	$V_{\mathrm{act}}$	81.8	88.0	79.4	33.2	70.6
Max	On	$V_{\mathrm{act}}$	78.4	86.8	78.2	27.8	67.8
Mean	Off	$V_{\mathrm{act}}$	73.8	85.2	71.4	27.0	64.4
Max	Off	$V_{\mathrm{act}}$	—	—	—	32.6	—
Mean	Off	$V_{\mathrm{latent}}$	62.2	84.4	67.2	25.8	59.9

Appendix

Table 4: Comparison of distillation teacher, success filter, and verifier (SmolVLA 5k, 10 adaptation episodes, evaluation seeds 10–59, 500 episodes per suite). Success rates (%) with (a) single-candidate generation and (b) weighted aggregation of eight candidates. “Mean” uses the weighted-average chunk and “Max” the highest-scoring candidate as the teacher. The main-method setting is “Mean teacher, filter on,

V_{\mathrm{act}}

.” The

V_{\mathrm{latent}}

rows also use

V_{\mathrm{latent}}

for candidate evaluation.

Evaluation	Teacher	Spatial	Object	Goal	Long	Avg.
Single candidate	Full chunk ( $H_{1}$ )	81.8	88.0	79.4	33.2	70.6
Single candidate	Executed steps ( $H_{2}$ )	79.0	92.0	84.2	41.0	74.1
$+$ Agg-8	Full chunk ( $H_{1}$ )	88.2	95.4	87.8	57.8	82.3
$+$ Agg-8	Executed steps ( $H_{2}$ )	90.4	99.0	92.6	64.2	86.6

Appendix

Table 5: Full-chunk teacher (

H_{1}=50

, main recipe) vs.

H_{2}

-only teacher (loss masked to the executed 10 steps). Success rates (%) on LIBERO (clean), SmolVLA 5k, seeds 10–59 (

n{=}500

per suite). The full-chunk rows are those of Table 1 .

$N$	Spatial	Object	Goal	Long	Avg.
1	57.6 / 73.8	80.4 / 85.2	61.0 / 71.4	17.4 / 27.0	54.1 / 64.4
2	68.0 / 76.0	82.6 / 89.0	70.6 / 76.4	26.0 / 37.6	61.8 / 69.8
4	74.6 / 79.0	89.8 / 93.4	81.8 / 81.2	36.6 / 43.6	70.7 / 74.3
8	80.6 / 83.0	94.4 / 94.6	86.8 / 82.4	44.0 / 55.0	76.4 / 78.8
16	84.0 / 86.0	92.8 / 96.4	90.6 / 86.2	56.2 / 62.0	80.9 / 82.7

Appendix

Table 6: Numbers underlying Figure 3(a) : success rates (%) of Agg-

N

/ distilled policy + Agg-

N

, SmolVLA 5k,

n{=}500

per cell.

Policy (Base)	Method	$K{=}2$	$K{=}5$	$K{=}10$	$K{=}20$	$K{=}50$
5k (17.4)	Distill	29.6	29.2	27.0	29.2	34.6
	Distill + Agg-8	52.6	52.8	55.0	53.6	54.6
	N-GRPO	—	—	25.6	—	29.6
	DSRL	—	—	19.4	—	18.8
10k (68.6)	Distill	75.0	70.4	73.8	75.0	75.4
	Distill + Agg-8	75.2	76.8	79.0	80.4	85.4

Appendix

Table 7: Success rate (%) on LIBERO-Long as a function of the adaptation-episode budget

K

(

n{=}500

). Parameters are frozen after

K

adaptation episodes and then evaluated. DSRL-style is our port of the official implementation (dsrl_pi0), with a

-1/0

reward (time penalty), a TanhNormal distribution restricted to

[-1,1]

, an ensemble of 10 critics, and prior-replacement composition. The GRPO rows are the comparison method (verifier-ranking reward, top1 advantage, success filter, argmax adaptation). The distilled policy is the variant without the success filter (Appendix B ), with adaptation seeds

0,\ldots,K-1

and the 50 seeds immediately following as evaluation seeds. Base performance is 17.4 at 5k and 68.6 at 10k (seeds 10–59); Agg-8is 44.0 at 5k and 80.0 at 10k.

Verifier	Method	Spatial	Object	Goal	Long	Avg.
—	Base	85.0	95.0	84.0	68.6	83.2
$V_{\mathrm{act}}$	Agg-8	90.2	99.6	91.8	80.0	90.4
	Distill	89.4	96.0	85.4	73.8	86.2
	Distill + Agg-8	91.8	98.4	89.8	79.0	89.8
$V_{\mathrm{latent}}$	Agg-8	92.4	99.0	91.4	82.6	91.4
	Distill	91.0	97.0	80.8	75.6	86.1

Appendix

Table 8: Success rates (%) on LIBERO (clean) with SmolVLA 10k. Measured under the same protocol as Table 1 (evaluation seeds 10–59,

n{=}500

per suite).

clean 5k
Method	Spatial	Object	Goal	Long	Avg.
Base	57.6	80.4	61.0	17.4	54.1
Agg-8 $V_{\mathrm{act}}$	80.6	94.4	86.8	44.0	76.4
Agg-8 $V_{\mathrm{latent}}$	74.0	89.8	81.2	41.6	71.7
$\Delta$	$-6.6\pm 3.7$	$-4.6\pm 1.7$	$-5.6\pm 2.5$	$-2.4\pm 2.2$	$-4.8\pm 1.3$
clean 5k, as distillation teacher

Appendix

Table 9: Comparison of verifier inputs (success rate %; Agg-8is softmax aggregation with

\lambda{=}0.3

; each verifier trained for 60k steps). Clean uses evaluation seeds 10–59 (

n{=}500

per suite); plus uses 105 tasks

\times

10 episodes (seeds 10–19).

\Delta

is the per-task paired difference

V_{\mathrm{latent}}-V_{\mathrm{act}}

(mean

\pm

SEM). The comparison as a distillation teacher is based on the variant without the success filter (

K{=}10

, single-candidate generation).

Figure 5: Scatter plot comparing Agg-8success rates on all 420 LIBERO-plus tasks (SmolVLA 10k): $V_{\mathrm{act}}$ (horizontal) vs. $V_{\mathrm{latent}}$ (vertical), 10 episodes per point (jitter of $\pm 1.5$ points added to reduce overlap). Points above the diagonal favor $V_{\mathrm{latent}}$ . The upward shift is especially pronounced on the Object suite.

Item	Feature-space comparison (Figure 6 )	Main experiments
Policy	BC-trained on the two suites Object and Long. 4k checkpoint ( run_bc_success/checkpoints/004000 )	BC-trained on successful rollouts from all four suites. 5k / 10k / 20k checkpoints of run_bc_4suite
Candidate data	cand_4k_a1.2 . Object and Long candidates generated from the 4k policy	cand_4suite_<ckpt>_a1.2 . Four suites, generated per policy checkpoint
Verifier training	20k steps	60k steps
World model	run1 (SmolVLA features) and run_dinov2 (DINOv2 features)	run_4suite (SmolVLA features, all four suites)
Candidate noise at evaluation	$\alpha=2.0$	$\alpha=1.2$
Execution rule	Select the highest-scoring of 8 candidates (argmax)	Softmax-weighted aggregation of 8 candidates ( $\lambda=0.3$ )

Appendix

Table 10: Differences in setup between the feature-space comparison (Figure 6 ) and the main experiments (clean LIBERO). “a1.2” in the candidate-dataset names denotes the noise multiplier.

Figure 6: Comparison of verifier input representations with a SmolVLA 4k checkpoint trained on the Object and Long suites. Eight candidates, noise scale $\alpha=2.0$ , argmax selection, 25 evaluation episodes per task (250 per suite). Verifiers are trained for 20k steps. The base policy, $V_{\mathrm{act}}$ with direct action input, and $V_{\mathrm{latent}}$ predicting DINOv2 features or SmolVLA features through the world model are compared. See Table 10 for details of the setup.

Checkpoint	Suite	Base	$\alpha{=}1.2$	$1.5$	$2.0$	$3.0$
5k	Object	82	82	59	39	44
5k	Long	17	42	21	7	6
10k	Object	96	98	93	64	71
10k	Long	67	74	60	27	21
20k	Object	99	98	91	84	75
20k	Long	90	79	68	33	30

Appendix

Table 11: Success rate as a function of the candidate noise scale

\alpha

(

N{=}8

V_{\mathrm{latent}}

n{=}100

per point, Long / Object suites). Bold marks the maximum in each row including the base policy (ties included). The explored range is

\{1.2,1.5,2.0,3.0\}

Checkpoint	$R$	Split-cos	SDR
5k	$0.229\pm 0.007$	$0.219\pm 0.025$	$0.053\pm 0.004$
10k	$0.228\pm 0.007$	$0.216\pm 0.027$	$0.053\pm 0.004$
20k	$0.209\pm 0.005$	$0.129\pm 0.023$	$0.043\pm 0.002$

Appendix

Table 12: Within-state reproducibility of the TTC correction vector

\Delta_{m}

(32 states, 32 candidate-set draws per state). (a) Directional reproducibility metrics; (b) correction norm, effective number of candidates, and score standard deviation.

R

is the resultant length, Split-cos the split-half cosine similarity, and SDR the directional stability ratio.

Method	Success	$\Delta$ Base	$\Delta$ Agg-8
Base	$37.3\pm 1.1$	—	$-6.6\pm 0.7$
Agg-8	$43.9\pm 1.2$	$+6.6\pm 0.7$	—
N-GRPO + Agg-8	$43.8\pm 1.2$	$+6.5\pm 0.7$	$-0.1\pm 0.6$

Appendix

Table 13: Success rates (%) on LIBERO-plus with SmolVLA 20k and per-task paired differences from Base (points, mean

\pm

SEM). Evaluation seeds 10–19. (a) Aggregated over the 837 Level-5 tasks on which Base has a positive success rate; (b) aggregated over a fixed set of 80 tasks (32 tasks for PRO only). VWFM is advantage-weighted flow matching over all candidates, Online is online LoRA updating, and PRO is preference optimization (Appendix M ).

Checkpoint	Method	Spatial	Object	Goal	Long
SmolVLA 5k	Base	62.6	80.0	59.6	17.4
	Sel-8	73.6	88.2	81.8	34.4
	N-GRPO	59.8	83.8	68.2	23.2
	N-GRPO+Sel-8	80.2	88.2	78.8	53.4
SmolVLA 10k	Base	85.0	96.4	84.0	65.6
	Sel-8	89.6	97.8	86.8	77.0

Appendix

Table 14: GRPO success rates (%) on clean LIBERO, 50 episodes per task. Base and Sel-8use evaluation seeds 0–49 (argmax selection); the GRPO rows use adaptation seeds 0–9 and evaluation seeds 10–59.

Method	Spatial	Object	Goal	Long	Avg.
N-GRPO ( $V_{\mathrm{act}}$ )	35.9	21.6	13.8	18.0	22.3
N-GRPO+Sel-8 ( $V_{\mathrm{act}}$ )	42.1	29.6	15.3	21.3	27.1
N-GRPO+Sel-8 ( $V_{\mathrm{latent}}$ )	42.4	32.5	13.4	16.7	26.3

Appendix

Table 15: GRPO success rates (%) on LIBERO-plus 10k, 105 tasks

\times

10 episodes, argmax selection (Sel-8). GRPO applies the success filter and the

\mu

trust region. Parentheses indicate the verifier used for the reward and for candidate selection.

Advantage scheme	Clean 5k Long	Plus Object $V_{\mathrm{act}}$	Plus Object $V_{\mathrm{latent}}$	Plus Long $V_{\mathrm{latent}}$
top1	53.4	29.6	32.5	16.7
zscore	52.6	30.0	30.8	16.4
rank	48.6	28.0	31.7	15.4
tophalf	48.4	27.4	33.0	14.7

Appendix

Table 16: Comparison of advantage computation schemes (

n{=}500

n{=}1050

). Assignment schemes that dilute the weight of the top candidates tend to weaken the correct update direction. Settings with

V_{\mathrm{latent}}

reward and the

\mu

trust region were also tested.

Explore similar work

VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model
May 2, 2026Wenhao Li, Xiu Su, Yichao Cao +5Vision-Language-Action Models Test-Time Adaptation of VLA Models
University of Sydney · Central South University · University of Science and Technology of China +2
TTT-VLA: Test-Time Latent Prompt Optimization for Vision-Language-Action Models
Jun 2, 2026Wenbo Zhang, Jianxiong Li, Shuai Yang +4Test-Time Adaptation Test-Time Adaptation of VLMs
ByteDance Seed · The University of Adelaide · Tsinghua University +3
TokenPB: Enabling VLA Action Self-Verification via VLM Token Probability Bucketing
Mar 18, 2026Chen Zhao, Zhuoran Wang, Haoyang Li +6Vision-Language-Action Models
School of Information, Renmin University of China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China · Beijing University of Posts and Telecommunications +1

May 2, 2026J/K move · Enter open · S save

VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model

Wenhao Li, Xiu Su, Yichao Cao, Hongyan Xu+4

University of Sydney · Central South University · University of Science and Technology of China+2

Vision-Language-Action Models Test-Time Adaptation of VLA Models VLMs for Robotics

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models have demonstrated remarkable capabilities and generalization in embodied manipulation. However, their decision-making relies on a fast, instinctive process that lacks deliberation. This strategy often leads to suboptimal or catastrophic actions when facing complex or ambiguous scenarios that require greater consideration. In this paper, we introduce VLA-ATTC, a framework that endows VLA models with adaptive test-time compute (TTC). VLA-ATTC employs an uncertainty-based ``cognitive clutch'' to dynamically transition from reflexive execution to a TTC deliberation phase when necessary. During TTC phase, a novel Relative Action Critic (RAC) model identifies the optimal action from generated candidates via pairwise comparisons. This relative mechanism replaces unstable absolute value estimation, significantly simplifying the learning objective. Furthermore, we introduce an efficient sampling strategy to amortize computational costs and an automated data pipeline that curates preference pairs without manual annotation. On the LIBERO-LONG benchmark, VLA-ATTC reduces the failure rate of the SOTA model PI0.5 by over 50%. We will open-source all the code and weights.

Abstract

Explore similar work

VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model

TTT-VLA: Test-Time Latent Prompt Optimization for Vision-Language-Action Models

TokenPB: Enabling VLA Action Self-Verification via VLM Token Probability Bucketing