acceptodds
Under review as a conference paper at ICLR 2027

SparkVLA: Stop-Aware Hierarchical VLA with Adaptive Action Chunking for Long Horizon Manipulation

13 views

Abstract

At every re-observation point in a hierarchical Vision-Language-Action (VLA) system, two interface decisions must be made. The first is when to terminate the current subtask. The second is how far to execute the proposed action chunk. These decisions are mutually dependent. The useful stopping point depends on the proposed executor behavior, while the useful execution length depends on subtask progress. Many architectures nevertheless evaluate them with separate decision pathways, which leaves an information asymmetry that neither module can overcome alone. We present SparkVLA, a stop-aware hierarchical VLA that implements this coupled decision by formulating both choices as a single ranking. In this ranking, Stop competes against every action-prefix length in a unified candidate set, and the system selects the highest-scoring option, avoiding a separately tuned probability threshold and requiring only offline ordinal preferences. An Anchor-Conditioned Context Encoding module caches a history-aware subtask anchor that encodes onset-state memory and goal semantics, and guides visual-token pruning toward task-relevant regions. A Stop-Aware Action-Prefix Selection} head scores all candidates via full self-attention at chunk boundaries for efficiency. On RoboCerebra, SparkVLA achieves 36.1% complete-task success and 47.12% transition completion, surpassing the official hierarchical baseline by 30.10 percentage points on complete-task success. Real-robot experiments on multi-step tasks further validate these gains on physical hardware.

Primary area:
applications to robotics, autonomy, planning
Keywords:
long horizon manipulation, VLA, adaptive action chunking

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_LrUZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SparkVLA, a hierarchical Vision-Language-Action framework for long-horizon manipulation that jointly models subtask termination and adaptive action-chunk execution. Instead of using a separate stop detector and action-horizon selector, SparkVLA introduces a unified candidate set in which STOP competes directly with all possible action-prefix lengths. The framework further incorporates a history-aware subtask anchor, anchor-guided visual token pruning, and a stop-aware ranking objective trained from offline subtask-boundary annotations. Experiments on RoboCerebra, LIBERO, and three real-world manipulation tasks show consistent improvements over standalone VLAs, hierarchical baselines, and matched adaptive-chunking methods.

Critical strengths

The paper addresses an important and practical interface problem in hierarchical VLA systems: subtask termination and execution horizon are naturally coupled, yet are often handled independently. The proposed unified ranking formulation is simple, intuitive, and easy to integrate into existing hierarchical systems. The experimental evaluation is relatively comprehensive, including matched comparisons with DEHP, ACH, AQC, and AAC, interface-level ablations, executor transfer, proposal-shift diagnostics, and real-robot experiments. In particular, the matched study shows that SparkVLA improves complete-task success from 30.1% with factorized heads to 36.1%, while also substantially reducing premature subtask transitions. The physical experiments further support the practical value of the method, with SparkVLA outperforming both the base policy and AAC with matched Stop across all three tasks.

Critical weaknesses

The main limitation is that the conceptual novelty is somewhat incremental: treating STOP as another candidate in the action-prefix ranking space is elegant, but the paper occasionally overstates this as resolving a dependency that independent modules fundamentally cannot handle. A sufficiently expressive factorized model could in principle expose both heads to the same action proposal and context, so the empirical result is better interpreted as evidence for a useful inductive bias rather than a fundamental architectural necessity. In addition, the ordinal supervision is derived from recorded completion boundaries rather than counterfactual outcomes of the newly proposed action prefixes. Its validity therefore depends on alignment between the recorded trajectory and the executor proposal, which the authors themselves acknowledge, and the proposal-shift experiments show gradual performance degradation as this alignment decreases.

Questions for authors

My main question concerns the robustness of the boundary-derived supervision. Since the preferred prefix length is inferred from the remaining distance to the recorded completion boundary rather than from the actual outcome of each proposed prefix, how sensitive is SparkVLA when the executor proposal deviates more substantially from the recorded trajectory, especially under larger executor shifts or out-of-distribution states?

3 comments

We thank the reviewer for recognizing the practical value of the shared decision interface and the matched and physical evaluations. We clarify the learning-bias contribution and present the proposal-shift and execution evidence addressing your main concern about boundary-derived supervision.

Question 1 and Critical Weakness 2: Robustness to proposal shifts and correction states

SparkVLA retains an advantage over joint categorical selection in each tested condition. We examine changes to the executor weights and correction states to assess sensitivity to both the proposal-generating model and the encountered states:

\hline \begin{array}{c}\textbf{Evaluation}\cr \textbf{condition}\end{array} & \begin{array}{c}\textbf{Categorical SR}\cr \textbf{(\\%)}\end{array} & \begin{array}{c}\textbf{SparkVLA SR}\cr \textbf{(\\%)}\end{array} & \begin{array}{c}\textbf{Gain (pp),}\cr \textbf{95\\% CI}\end{array} \cr \hline \text{Standard} & 33.5 & \mathbf{36.1} & \begin{array}{c}+2.61\cr [0.72, 4.50]\end{array} \cr \begin{array}{l}\text{Same architecture,}\cr \text{different executor weights}\end{array} & 31.3 & \mathbf{34.3} & \begin{array}{c}+3.00\cr [0.83, 5.17]\end{array} \cr \text{OOD correction states} & 27.8 & \mathbf{31.7} & \begin{array}{c}+3.83\cr [1.61, 6.06]\end{array} \cr \hline \end{array}$$ Gains use unrounded rates and paired intervals. The ordinal gains are +3.00 pp for changed executor weights and +3.83 pp for correction states, with paired confidence intervals above zero in both conditions. Correction states include contact recovery, slip handling, and obstacle avoidance. We additionally evaluate decision-outcome agreement on 2,000 held-out (state, proposal) pairs. Agreement compares predicted accept/reject decisions with the observed execution outcomes: $$\begin{array}{lrrr} \hline \textbf{Audit group} & \textbf{Pairs} & \begin{array}{c}\textbf{Joint categorical}\cr \textbf{(\\%)}\end{array} & \begin{array}{c}\textbf{SparkVLA}\cr \textbf{(\\%)}\end{array} \cr \hline \text{Accepted proposals} & \text{1,300} & 68 & \mathbf{90} \cr \text{Rejected proposals} & 500 & 65 & \mathbf{84} \cr \begin{array}{l}\text{OOD correction-state}\cr \text{proposals}\end{array} & 200 & 59 & \mathbf{81} \cr \hline \end{array}$$ SparkVLA improves both complete-task success in the shift evaluations and decision-outcome agreement in all three audit groups. Section 4.7 and Appendices A.2–A.3 now report the evaluation conditions, paired intervals, and executed-outcome measurements. ## Critical Weakness 1: Contribution and the role of factorization **The matched controls quantify the benefits of the shared decision interface and stop-aware ordinal training:** joint selection raises task success from 30.1% to 33.5%, and the full ordinal selector reaches 36.1%. The factorized control receives the same proposal and context as the joint selectors, with matched model capacity and training cases. The design makes STOP and continuation lengths directly comparable in one candidate space. Within this space, ordinal training uses the same boundary to supervise relative preferences among valid prefixes, including non-optimal candidates. This distinguishes the choice of interface from the choice of training objective. The data-volume ablation further shows an ordinal advantage of +3.17 pp at 50% training data and +3.78 pp at 25%. The matched comparisons demonstrate the benefits of the shared STOP–prefix decision space and stop-aware ordinal training. We have emphasized these contributions in the abstract and Sections 1 and 3.2. Appendix A.1 explains the matched comparisons and reports the data-volume ablation in Table 6.

Thank you for the detailed response; you have addressed most of my concerns, and I am happy to support acceptance of your paper.

We sincerely appreciate your recognition of our work and your support for acceptance! Your thoughtful comments have helped us clarify our presentation and strengthen the manuscript. Thank you for taking the time to read our responses and for the constructive discussion throughout the review process.

Reviewer_PUSfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose SparkVLA, a hierarchical VLA framework that jointly determines subtask termination and action execution length. Instead of predicting STOP and action chunk length separately, the method formulates both decisions as a unified ranking problem. The framework further introduces a history-aware subtask anchor, adaptive visual token pruning, and a scoring head trained with offline ordinal preferences. Experiments on RoboCerebra and LIBERO show improvements over existing hierarchical VLA and adaptive chunking methods. Real robot experiments further demonstrate improved success rates on several long-horizon manipulation tasks.

Critical strengths

The paper addresses a meaningful problem in long-horizon robotic manipulation. I appreciate the idea of considering subtask termination and action chunk selection together, since both decisions affect the responsiveness and reliability of a hierarchical VLA system. The experimental evaluation is also relatively comprehensive. The authors compare against several recent adaptive chunking methods, including DEHP, ACH, AQC, and AAC, and provide additional controlled comparisons between separate and unified decision heads. The real robot experiments with 50 trials per task are another positive aspect of the work.

Critical weaknesses

My main concern is the limited novelty compared with the rapidly growing literature on adaptive VLA execution. Several recent papers have already explored dynamically adjusting computation or execution decisions according to manipulation stages. For example, AAC, DEHP, ACH, and AQC investigate adaptive action chunk selection using different signals and objectives. VLA-ADP (ICLR 2026) also explores action-aware adaptation across coarse and fine manipulation stages, although it focuses on visual token pruning rather than execution horizon selection. The authors propose a unified ranking formulation, but I find the technical contribution somewhat incremental given these existing directions. The main novelty appears to be adding STOP as another candidate in an adaptive action selection framework, rather than introducing a fundamentally new mechanism. Although the authors provide matched comparisons, Table 7 shows that a joint categorical baseline already achieves 33.5% success, compared with 36.1% for SparkVLA. This relatively modest improvement makes me question how much of the reported gain comes specifically from the proposed ranking formulation. Another concern is the supervision strategy. The preference labels are derived mainly from recorded subtask completion boundaries rather than the actual outcomes of executing different candidate prefixes. The authors acknowledge this limitation and provide proposal alignment diagnostics, but it remains unclear how reliably the ranking generalizes when generated actions deviate from recorded trajectories. Finally, the efficiency motivation could be better supported. In the physical experiments, SparkVLA has a decision latency of 290 ms compared with 260 ms for the original π0.5 baseline, so the practical computation overhead deserves further discussion. Overall, I find the experimental results encouraging, but the methodological novelty and the advantages over simpler adaptive decision mechanisms are not fully convincing.

Questions for authors

Could the authors clarify the fundamental technical difference between SparkVLA and existing adaptive chunk selection methods such as AAC, DEHP, ACH, and AQC, beyond introducing STOP into the candidate set? The paper already provides comparisons with these methods, but I would like to understand what new principle the unified ranking formulation establishes. Could the authors also discuss the relationship between their anchor-guided visual pruning and previous action-aware adaptive pruning methods such as ADP? Regarding the experiments, Table 7 shows a relatively small improvement over joint categorical selection. Is this difference consistent across tasks and statistically significant? I am also concerned that the ranking targets are constructed from recorded completion boundaries rather than counterfactual execution outcomes. How does this supervision behave when the proposed action trajectory differs substantially from the recorded one? Finally, could the authors report the overall computation and execution-time tradeoff, especially given that the real robot decision latency is higher than that of the original π0.5 baseline?

2 comments

Thank you for the constructive review. We have clarified the contribution in Sections 1 and 3.2, added the ADP comparison to Section 2.2, and expanded the supporting evidence in Appendices A and J.

Critical Weakness 1 and Question 1. Contribution beyond adaptive action chunking

SparkVLA jointly decides whether to commit the current subtask or how long to execute the current proposal before observing again. STOP and continuation lengths compete in one scoring space. Ordinal supervision specifies relative preferences among continuation lengths as well as between STOP and continuation.

AAC uses action entropy, DEHP learns an execution-horizon branch, and ACH and AQC use value-based horizon selection. Their horizon-selection objectives do not explicitly rank subtask completion against continuation. Our matched evaluations augment these methods with the same STOP controller.

The interface comparison holds context, proposal inputs, training cases, and model capacity fixed.

\hline \textbf{Selector} & \textbf{Task SR (\\%)} & \textbf{Early advance (\\%)} \cr \hline \text{Factorized STOP and horizon} & 30.1 & 6.00 \cr \text{Joint categorical} & 33.5 & 4.51 \cr \text{Joint ordinal (SparkVLA)} & 36.1 & 1.99 \cr \hline \end{array}$$ Early advance counts subtask commits before annotated completion. Moving from factorized to joint categorical selection adds 3.39 pp, with a paired 95% CI of [1.50, 5.28]. Moving from categorical to SparkVLA's stop-aware ordinal objective adds another 2.61 pp. Gains use unrounded rates. Categorical training identifies a preferred class. Ordinal training also supervises preferences between non-optimal continuation lengths according to their relation to the recorded completion boundary. These preferences reuse the same annotation. SparkVLA combines this ranking with a STOP auxiliary loss over the same scores. **Matched objective controls isolate the ranking contribution.** In the separate 350,000-comparison study, ordinal training outperforms categorical training under both matched STOP-auxiliary settings, by 2.1 pp without the auxiliary and 3.2 pp with it. Removing only prefix-prefix ranking while retaining the STOP-related losses and their coefficients lowers task success from 35.7% to 33.6%. Restoring these terms adds 2.1 pp, with a 95% CI of [0.90, 3.33] (Appendix J.2). The original all-retained-data model remains 36.1%. The ordinal selector also reduces median completion-to-commit delay from 275 to 225 ms relative to categorical selection (Appendix F.2, Table 11). Together, these results support joint comparison and relative-length supervision as useful design choices beyond adding STOP as a candidate. Section 4.4 documents the matched STOP wrapper. ## Question 2. Relationship to action-aware visual pruning ADP and SparkVLA use different conditioning signals for a related visual-efficiency objective. ADP adjusts token retention using recent action trajectories and text-driven selection. SparkVLA uses a cached subtask anchor incorporating onset information and subtask history. Its visual sampling supplies subtask-conditioned features for the central STOP-execution-length decision. We have added the ADP citation and this comparison to Section 2.2. ## Critical Weakness 1 and Question 3. Size and consistency of the gain **SparkVLA achieves 47 additional complete-task successes in 1,800 episodes**, increasing successes from 603 to 650 relative to joint categorical selection. The gain is 2.61 pp, with a paired 95% CI of [0.72, 4.50], using 10,000 case-paired bootstrap resamples. The one-sided p-value is 0.004. All three reported run-wise gains are positive, at +3.83, +2.33, and +1.67 pp. Task-level examples span drawer opening (18/30 to 24/30), bowl placement (9/30 to 15/30), rack assembly (12/30 to 16/30), and stove interaction (8/30 to 11/30), listed as categorical versus ordinal. The advantage also persists at 50% and 25% training data, with gains of 3.17 pp (95% CI [1.00, 5.33]) and 3.78 pp ([1.50, 6.06]), respectively. Appendix A.1 and Table 6 report these analyses, and Appendix F.7 gives the per-run success counts.

Critical Weakness 2 and Question 4. Supervision under proposal deviation

SparkVLA maintains its task-success advantage under proposal shifts and makes more accurate relative-prefix choices than categorical selection in the execution audit. With different weights of the same executor architecture, SparkVLA achieves 34.3% task success versus 31.3% for categorical selection. On correction states, the rates are 31.7% versus 27.8% (Section 4.7 and Appendix A.2).

To directly examine the boundary-derived ordering, we execute alternative lengths of the same proposal from the same restored state. The audit uses 360 held-out snapshots from 60 cases, with 120 snapshots each in Standard, Lower-alignment, and Correction cohorts. Each branch executes its prefix, re-observes, and follows the same frozen continuation policy under a common deadline. Tested lengths are 1, 5, 10, 20, 35, and 50 actions.

Independent discovery and confirmation batches each use 50 rollouts per state-length branch. Discovery identifies outcome-resolved pairs, and the same set is evaluated for all methods against independent confirmation outcomes. Coverage is the fraction of scheduled prefix pairs entering this shared evaluation. Agreement scores matching directions as 1, opposite directions as 0, and ties as 0.5. Ordering diagnostics use one predesignated matched seed. Appendix J.3 gives the full protocol.

\hline \textbf{Cohort} & \textbf{Coverage (\\%)} & \textbf{Boundary (\\%)} & \textbf{Categorical (\\%)} & \textbf{SparkVLA (\\%)} \cr \hline \text{Standard} & 80.0 & 91.0 & 68.0 & 85.0 \cr \text{Lower-alignment} & 70.0 & 83.0 & 62.0 & 78.0 \cr \text{Correction} & 65.0 & 76.0 & 58.0 & 77.0 \cr \hline \end{array}$$ Boundary-derived preferences agree with execution outcomes on 83.0% of evaluated pairs in the Lower-alignment cohort and 76.0% in Correction. SparkVLA's predicted ordering agrees with outcomes more often than categorical selection in every cohort. The correction-state difference is 19.0 pp, with a paired 95% CI of [10.72, 27.09]. These measurements directly test relative execution-length choices. We also examine cases where the recorded boundary favors a longer prefix but execution evidence favors earlier re-observation. On 200 discovery-selected conflict pairs, comprising 50 Standard and 150 Correction pairs, SparkVLA prefers the shorter prefix on 80.0% of pairs, versus 65.0% for categorical training. Independent confirmation gives 78.0% branch completion after the shorter prefix versus 60.0% after the longer prefix under the same continuation policy and deadline (Appendix J.4). This shows that the learned selector can favor earlier feedback when a boundary-preferred continuation is less effective. The action-content control provides complementary evidence. Full-proposal encoding reaches 36.1% task success, compared with 31.1% for action-blind and 33.3% for terminal-only encoding under the same boundary labels and parameter budget (Appendix F.6, Table 15). The selector therefore uses the proposed action content alongside current context, rather than relying only on a boundary-based stopping rule. ## Critical Weakness 3 and Question 5. Overall computation and execution cost **SparkVLA improves physical-task success and reduces mean whole-trial time while adding 30 ms to decision p95 relative to $\pi_{0.5}$.** Each method receives 150 physical trials under the same task deadlines and reset rules. Wall time includes failures, waiting, and planner overhead. $$\begin{array}{lrrr} \hline \textbf{Method} & \textbf{Successes / trials} & \textbf{Mean wall time (s)} & \textbf{Decision p95 (ms)} \cr \hline \pi_{0.5} & 72/150 & 244.56 & 260 \cr \text{AAC + matched STOP} & 92/150 & 221.12 & 420 \cr \text{SparkVLA} & 104/150 & 202.20 & 290 \cr \hline \end{array}$$ Relative to $\pi_{0.5}$, decision p95 rises from 260 to 290 ms, while mean wall time falls by 17.3% and success rises by 21.3 pp. Against AAC with matched STOP, success improves by 8.0 pp and mean wall time falls by 18.92 s. These measurements quantify the practical tradeoff between decision overhead, whole-trial duration, and task completion. Section 4.6 now reports whole-trial time, and Appendix A.4 presents the success-time-latency comparison.
Reviewer_Zfc2OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

SparkVLA jointly ranks STOP and candidate action-prefix lengths within a hierarchical VLA, using proposal-aware representations, ordinal supervision and an auxiliary stopping loss. It also incorporates subtask history, an onset anchor and efficient visual sampling. On RoboCerebra, the full selector reaches 36.1% task success versus 30.1% for a matched factorized selector.

Critical strengths
  1. Treating subtask stopping and action-prefix selection as one decision is a clear and useful formulation. It addresses the interaction between execution length and subtask transitions instead of optimizing the two decisions independently.
  2. The controls are relatively strong. Tables 3, 7, 11, and 12 include matched adaptive-horizon methods, factorized and categorical selectors, loss/input ablations, and three selector seeds. The full model's 36.1 ± 0.1% versus 30.1 ± 0.9% for the factorized selector supports the joint formulation under this setup.
  3. The real-robot evaluation uses 50 interleaved trials per task and reports both task success and latency. Proposal-shift and action-blind controls also acknowledge and investigate the limitations of demonstration-derived supervision.
Critical weaknesses
  1. The proposal-to-supervision alignment procedure needs an operational specification. Demonstration boundaries do not automatically provide counterfactual labels for a newly sampled action proposal, as the paper acknowledges. Appendix E.7 refers to proposal matching, retained-pair fractions, and paired-bootstrap intervals, but the matching criterion, actual retained fractions, and those intervals are not clearly reported.
  2. Several numerical inconsistencies should be corrected. Table 1 gives Mem-0 task success of 9.49%, so the gain to 36.1% is 26.61 points, not the 27.54 stated in Section 4.2. The four π0.5 entries in Table 2 average to 96.95%, while the table reports 96.85%; Appendix Table 10 lists a different set. The default timeout rate is 2.28 in Table 3 but 2.78 for L = 2 in Table 9.
  3. The ordinal supervision in Eq. (13) encodes a strong temporal preference: before reaching the recorded completion boundary, longer prefixes receive higher priority. Although this does not require physically monotonic trajectories, it may overlook the value of early re-observation during retries, backtracking, or contact-rich corrections.
Questions for authors
  1. Please provide the exact proposal-matching rule, retained-pair fractions, paired-bootstrap intervals, and training-data/annotation counts for all selector controls.
  2. Can you reconcile the Mem-0 gain, π0.5 LIBERO entries, and default timeout rate, and distinguish task success from task score consistently in figures and tables?
  3. How does performance change as proposal matching becomes looser when the retained training-data budget is held fixed? This would help distinguish label fidelity from sample-count effects.
1 comment

We thank the reviewer for recognizing the joint formulation, matched controls, and physical evaluation. We have clarified proposal alignment and supervision, corrected the numerical inconsistencies, and distinguished complete-task success from transition completion.

Question 1 and Critical Weakness 1: Proposal alignment and supervision

The matching rule retains recorded comparisons within 1.0 s and 5 cm, with matching gripper states. At this setting, 80% of boundary states have valid comparisons. Training uses 600,000 pairwise comparisons derived from recorded completion-boundary annotations. Record-end validity and subtask-status masks determine which candidates enter the loss (Appendix B).

The interface comparison holds proposal inputs, context, training cases, and model capacity fixed. Joint categorical selection gains +3.39 pp over factorized heads (95% CI [1.50, 5.28]); the stop-aware ordinal selector adds +2.61 pp ([0.72, 4.50]), corresponding to 650/1,800 versus 603/1,800 complete-task successes. The latter interval uses 10,000 case-paired bootstrap resamples.

Boundary matching supplies a temporal preference for recorded behavior; execution agreement separately checks decisions against observed outcomes. We have clarified the paired analysis in Appendix A.1 and the matching thresholds, coverage, used training counts, and execution audit in Appendix A.3.

Question 2 and Critical Weakness 2: Numerical corrections and metric definitions

We have corrected the Mem-0 gain, public π0.5\pi_{0.5} LIBERO entries, and default timeout rate. The Mem-0 gain is 36.1−9.49=26.6136.1-9.49=26.61 pp. The public LIBERO row is 98.8/98.2/98.0/92.4, averaging 96.85%, consistently in the main and appendix tables.

The default uses L=2L=2 consecutive fresh observations for STOP confirmation, with 2.28% timeout; the 2.78% entry was a typo. Task SR denotes complete-episode success, and TS denotes transition completion: Table 1's 36.1% is task SR, whereas 47.12% and its condition columns report TS. Decision p95 and completion-to-commit delay are labeled separately.

The numerical corrections appear in Section 4.2, Table 2, and Appendix F.4. We have also reconciled Table 2 with the public LIBERO row in Appendix F.5; Section 4.1 and the Table 1 caption give the metric definitions.

Question 3: Sensitivity to looser matching

Task success ranges from 34.8% to 36.7% across the matching settings below. Each setting uses all retained pairs for training, with training size varying from 350k to 780k pairs.

\hline \begin{array}{c}\textbf{Tolerance}\cr \textbf{(s)}\end{array} & \begin{array}{c}\textbf{Distance}\cr \textbf{(cm)}\end{array} & \textbf{Used pairs} & \begin{array}{c}\textbf{Task SR}\cr \textbf{(\\%)}\end{array} \cr \hline 0.5 & 3 & 350\mathrm{k} & 36.7 \cr 0.5 & 5 & 380\mathrm{k} & 36.4 \cr 1.0 & 5 & 600\mathrm{k} & 36.1 \cr 2.0 & 5 & 720\mathrm{k} & 35.2 \cr 2.0 & 10 & 780\mathrm{k} & 34.8 \cr \hline \end{array}$$ From the strictest to the loosest setting, coverage increases from 62% to 92%, while observed task success changes from 36.7% to 34.8%. The reported confidence intervals for the success-rate differences include zero. Appendix A.3 now reports the thresholds, coverage, and actual training counts together. ## Critical Weakness 3: Temporal preference and early re-observation **On correction states involving contact recovery and slip handling, SparkVLA reaches 31.7% task success versus 27.8% for joint categorical selection**, a +3.83 pp gain with a paired 95% CI of [1.61, 6.06]. The supervision target is temporal progress toward the recorded completion boundary. At execution time, the selector scores STOP and valid action prefixes using the current context and action proposal; the selected prefix length determines when the next observation is acquired. Section 3.4 specifies the temporal target, and Appendix A.2 reports performance in these correction settings.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.