acceptodds
Under review as a conference paper at ICLR 2027

Can Internal States Signal Trouble and Guide Recovery in VLAs?

Abstract

Frozen vision-language-action (VLA) policies can keep acting after progress has stalled. We introduce NAR, a test-time harness around a frozen mixture-of-experts VLA, driven by the expert routing already computed during action generation. On LIBERO-10, HiMoE-VLA failures exhibit stronger routing recurrence despite similar numbers of experts used per inference. NAR measures recurrence relative to the episode’s opening and detects deviations from successful source executions using nearest-neighbour scoring, temporal smoothing, and episode-level conformal calibration. Alarms trigger a fixed release-and-lift primitive before control returns to the same policy, requiring no parameter optimization or additional policy forward passes. Under a shared monitoring pipeline, routing achieves higher recall than the evaluated state and action readouts at their calibrated operating points. In a retrospective cross-suite evaluation, a profile fitted on native and perturbed LIBERO-Goal successes detects 97.2% of failures at a 1.32% episode-level false-alarm rate across 4,320 episodes from the other three suites, without target-task reference trajectories. With a separate pooled profile, paired replay of 1,608 completed alarms yields 175 net additional successes: 10.88 percentage points per alarm and 0.40 points when normalized by the recorded corpus. A separate timing study finds the lowest observed harm at the alarm, while later interventions lose opportunities to act. Transfer depends on source-success coverage, with specificity degrading on CALVIN. These results show how a policy’s existing routing trace can support a test-time harness that connects execution monitoring with the measured benefits and harms of intervention.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Embodied AI, Internal-State Dynamics, Execution Monitoring, Failure Detection, Recovery Control, Online Risk Assessment, Test-Time Intervention, Feedback-Guided Handoff, Frozen Policies, Robot Learning, Neural Alarm and Recovery

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9yB2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces a monitor for a frozen mixture-of-experts (MoE) VLA model based on its expert-routing patterns. The monitor tracks changes in these patterns during execution. When the patterns suggest that the policy may be stuck, it raises an alarm and triggers a recovery action.

Critical strengths

1.The paper examines its expert-routing patterns to decide intervention, without retraining the policy or requiring extra policy forward passes. Thus, the proposed method is general and easy to apply. 2. It evaluates whether responding to an alarm actually helps, accounting for both failures that are rescued and otherwise successful runs that are disrupted.

Critical weaknesses
  1. The proposed method is largely an engineering solution, and the paper reads more like a technical report than a research paper presenting a novel contribution.
  2. The intervention is triggered by observed expert-routing patterns and does not require gradients from the VLA model. However, the paper offers limited theoretical justification for why these patterns should indicate when intervention is beneficial. The intervention can both rescue failing executions and disrupt otherwise successful ones, so the evidence for its effectiveness is primarily empirical.
  3. The intervention is essentially a simple, reset-like action triggered by observed expert-routing patterns. Its practical utility is therefore unclear, and a more adaptive recovery strategy would make the contribution more compelling.
  4. The model’s internal routing is difficult to interpret, the paper should provide stronger evidence that this association is robust rather than coincidental.
  5. The proposed method is evaluated only in simulation, which limits its practical significance.
  6. The paper should describe its statistical analysis and explain how paired continuations were generated from the same simulator state to ensure a fair comparison. Potential failure cases should also be discussed more carefully.
Questions for authors

See weakness.

4 comments

Thank you for recognizing the routing interface that requires neither policy retraining nor additional policy forward passes, and the joint evaluation of recovery benefits and disruption costs. We clarify the contribution, intervention rationale, dynamical evidence, and evaluation protocol below.

W1: Research contribution and novelty

The contribution is to turn the internal routing dynamics of frozen MoE VLAs into a monitoring signal transferable across tasks, and to evaluate its practical control value through interventions paired at the same states. In frozen HiMoE-VLA [1], NAR reads routing already produced during normal action generation, without policy retraining or additional policy forward passes (§3).

The research has three connected components. First, we characterize differences in the temporal routing structure of successful and failed executions, and test their monitoring value through comparisons of observation signals and temporal statistics (§5.2; Tables 1 and 3). Second, we investigate transfer of successful reference dynamics: in Table 2(a), monitoring F1 is 82.7% on unseen initial states, 87.6% on held-out tasks, and 84.2% on a held-out task suite; Table 2(c) further evaluates LIBERO-Plus perturbations. Each suite uses its own frozen policy checkpoint; the monitoring representation and successful-source references are transferred, with native and perturbed successes included in the reference data. Third, paired continuations connect recovery benefits to the cost of disrupted successes, testing whether the internal signal can help improve task outcomes (§4; §5.4; Table 11).

These studies provide empirical evidence for understanding the relationship between internal computation and execution behavior in MoE VLAs, together with a concrete monitoring and intervention approach for improving task execution while keeping the model frozen.

W2/W3: Fixed recovery, net utility, and successful execution quality

The release-and-lift primitive changes local contact and end-effector conditions before returning control to the same frozen policy. Rather than resetting the environment, it may create an opportunity for learned approach and grasp skills to resume (§4; Appendix E.3).

Using the rescue–harm decomposition in §4, Eq. (2), the 1,608 state-matched pairs yield 285 rescues and 110 harms, or 175 net additional successes (+10.88 pp per completed alarm, 95% CI [4.67, 16.83]; §5.4; Table 11).

Eventual success need not imply smooth execution. In §5.2, 27 of 41 alarms on ultimately successful runs correspond to stalls followed by spontaneous recovery. In supplementary analysis, four successful but alarmed episodes require 27.6%–49.2% more actions than the median of their same-task successful controls; two exhibit sustained commanded motion with insufficient actual displacement before the alarm. Thus, a success label can conceal execution difficulty. Harms from the fixed intervention also show that recovery must respect opportunities for spontaneous recovery and the current manipulation state; all harms remain counted in our net estimate. Other response explorations are in Table 14.

W4: Why routing recurrence relates to execution difficulty

NAR does not equate repeated expert selection with failure. It compares recent routing dynamics with the episode's opening and evaluates deviations from successful-reference dynamics (§3; Eq. (1); Appendix A). In a perception–action loop, repeated unsuccessful interaction under similar conditions can induce persistent computational recurrence; ordinary expert reuse is also present in successful executions.

Three existing controls support this interpretation: (i) successes and failures use similar numbers of experts per inference (27.59 vs. 27.32), yet differ in temporal evolution (§5.2; Fig. 3); (ii) at a matched 55 false alarms, lag-three contraction achieves 94.3% recall, compared with 58.2% for adjacent-inference routing change (descriptive; Table 3a); and (iii) episode-consistent expert renaming retains 96.81% recall, whereas independently renaming experts at each inference breaks temporal correspondence and reduces recall to ≤0.17% without recalibration (§5.3; Fig. 6).

Together, these results support routing temporal dynamics as an internal observable of closed-loop execution difficulty, rather than a simple count of repeated experts.

References

[1] Du et al. HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies. https://arxiv.org/abs/2512.05693

W5: Real-robot applicability and broader internal signals

We thank the reviewer for highlighting real-robot risks from release and lift. Simulation allows state-matched intervention comparisons, while NAR itself only reads routing produced by the policy, without simulator ground-truth inputs (§3; Appendix D.5). On hardware, the current grasp and collision conditions should determine whether release or lift is permitted; recovery actions must satisfy robot-specific safety constraints.

Monitoring internal dynamics through neuron activations offers a path to extending this idea to dense models. NAD (Neuron Agreement Decoding) selects LLM responses based on agreement in neuron activations, while RAD (Routing Agreement Decoding) uses agreement in MoE routing to select candidates [3,4]. Both motivate investigating internal signals for execution monitoring and recovery in dense and MoE embodied policies.

Advances in MoE LLMs, together with embodied architectures such as HiMoE-VLA and AdaMoE [1,2], suggest that sparse computation will be an important direction for scaling future embodied models, further motivating our study of runtime monitoring for frozen MoE policies.

W6: Fair paired continuations, statistical analysis, and failure cases

The paired protocol and statistical analysis are detailed in §4, §5.1, and Appendices D.3 and D.5, including prefix replay, camera/state checks, same-state branches with shared noise, and configuration-level bootstrapping. The paper also quantifies three adverse patterns: disruption of otherwise successful continuations (Table 11), lower rescue rates with little budget remaining (Table 12), and repeated recovery crowding out policy actions (Table 4; Appendix E.3).

References

[1] Du et al. HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies. https://arxiv.org/abs/2512.05693

[2] Shen et al. Expertise need not monopolize: Action-Specialized Mixture of Experts for Vision-Language-Action Learning (AdaMoE). https://arxiv.org/abs/2510.14300

[3] Chen et al. Do LLMs Signal When They're Right? Evidence from Neuron Agreement (NAD). https://arxiv.org/abs/2510.26277

[4] Chen et al. Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control (RAD). https://arxiv.org/abs/2606.22798

I thank the reviewers for their detailed responses. It address most of my concerns, so I have decided to raise the score by one point. I will respond to each point individually to explain how the responses address my concerns.

W1/W2/W3: The clarifications provide a better understanding of the contributions. I recommend that the authors revise the abstract and introduction to highlight the research problem more clearly, as the current writing style is somewhat too empirical. Some concerns remain regarding the connection between the MoE architecture and the proposed interventions beyond the empirical results, as well as the generalizability of the method. I also encourage the authors to include a more detailed discussion of failure cases in the introduction and provide more insights than results. Although the interventions are simple, I appreciate their simplicity and effectiveness. However, the work could be more meaningful if it went beyond empirical evidence. W4: Addressed. W5: The explanations provided are acceptable. W6: Addressed.

Thank you for your detailed follow-up and for updating your assessment. We are glad that our responses have addressed most of your concerns, and we sincerely appreciate your careful consideration and constructive feedback on our work.

Reviewer_31LMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NAR (Neural-Alarm-Recovery), a test-time monitoring harness for frozen mixture-of-experts (MoE) vision-language-action (VLA) policies. The core observation is that failing executions exhibit stronger routing recurrence — recent inferences repeatedly select the same experts — compared to successful ones, despite using a similar number of experts per inference. NAR exploits this signal without any additional sensors, policy forward passes, or failure labels: it reads the expert routing trace already computed during action generation, computes a recurrence descriptor, and fires an alarm via split-conformal calibration when the descriptor deviates from a bank of successful reference episodes. The paper's secondary but arguably more durable contribution is a rigorous framework for evaluating whether acting on an alarm is worth it. By decomposing outcomes into rescue rate, harm rate, and break-even precision π⋆, the authors move beyond detection accuracy to quantify the net benefit of intervention. Experiments on LIBERO show that routing achieves higher failure recall than state and action baselines under a shared pipeline, and a Goal-suite profile transfers to 4,320 unseen episodes at 97.2% recall with 1.32% false-alarm rate. Paired replay of 1,608 alarms yields 175 net additional successes with a simple release-and-lift primitive, though the normalized gain across the full corpus is modest at 0.40 percentage points.

Critical strengths
  1. Architecturally grounded signal with zero marginal cost. Leveraging expert routing as a monitoring signal is an elegant match for MoE VLAs: the routing trace is an inherent byproduct of the forward pass already computed during action generation, requiring no extra sensors, hidden-state probes, or additional inference steps. The reported 3.14 ms per-update CPU overhead makes this approach genuinely practical for real-time deployment.

  2. Success-only calibration with formal false-alarm guarantees. NAR requires zero failure data during training—a crucial realism constraint for real-world deployment. By employing split-conformal calibration, the method delivers a principled, distribution-free bound on episode-level false alarms under exchangeability, offering a much stronger theoretical guarantee than the ad-hoc threshold tuning prevalent in prior literature.

  3. A principled evaluation framework for interventions. The rescue-and-harm decomposition, along with the concept of break-even precision, constitutes the paper’s most enduring contribution. It successfully shifts the diagnostic narrative from "does the alarm trigger correctly?" to "does acting on the alarm yield a net positive in expectation?"—a vital distinction largely overlooked by the community. Crucially, this framework is modular and applies independently of whatever signal drives the alarm.

  4. Transparent cross-suite transfer with well-defined boundaries. While the Goal profile scaling to 4,320 episodes across three unseen suites with 97.2% recall is an impressive empirical feat, the authors deserve equal credit for their transparency. They candidly acknowledge that specificity degrades on CALVIN and that conformal guarantees do not hold under distribution shift. Such self-aware academic honesty is rare and significantly bolsters the paper's credibility.

  5. Methodologically sound paired-branch evaluation. Branching off from state-matched replay points, rather than running entirely independent rollouts, is the correct way to isolate the efficacy of interventions from the noise of stochastic environment divergence. Furthermore, the explicit reporting of both rescue and harm rates—instead of hiding behind a net success lift—exemplifies rigorous and transparent experimental reporting.

Critical weaknesses
  1. Core novelty claim is unsubstantiated against the closest prior work. VLA-FAIL (Seligmann et al., 2026)—which the authors themselves acknowledge as being "closest to us"—also fits a success-only detector on internal VLA features using conformal calibration, evaluated on the exact same class of frozen policies. The sole distinguishing feature is the choice of monitoring signal: last-layer Mahalanobis distance versus routing recurrence. Yet, the paper provides zero direct empirical comparison between the two. The baselines in Table 1 simply adapt NAR's own pipeline to alternative input signals (such as state, action, and proprioception), which is fundamentally different from running VLA-FAIL. Consequently, the claim that routing features outperform last-layer representations rests entirely on assertion rather than evidence—precisely where the burden of proof is highest.

  2. Narrow evaluation scope: a single architecture and a single failure mode. All experiments are conducted using a single model (HiMoE-VLA) and restricted to a single failure type (timeouts). Routing recurrence is inherently coupled with the "stuck-in-a-loop" failure pattern characteristic of timeouts, which leaves open the crucial question of whether this signal can generalize to other failure modes, such as early catastrophic drops or failed grasps. More critically, the proposed method is, by construction, inapplicable to dense VLAs that lack discrete routing mechanisms—a substantial and rapidly growing fraction of modern deployed policies. While the authors briefly acknowledge this limitation in the conclusion, they offer no discussion on what an analogous continuous-domain signal might entail, leaving the actual scope of the contribution considerably narrower than what the title implies.

Questions for authors

Q1. The paper identifies VLA-FAIL as the closest prior work, yet the empirical comparison remains indirect: NAR's pipeline is evaluated on alternative signals rather than running VLA-FAIL's Mahalanobis detector directly on the same episodes. Could the authors provide a direct head-to-head comparison? Alternatively, if VLA-FAIL's implementation is unavailable for rebuttal, could you offer a principled rationale for why routing recurrence should theoretically outperform last-layer features on this task? Without empirical or theoretical backing here, the primary novelty claim remains unsupported at the exact point where the burden of proof is highest.

Q2. All 1,373 failures in the corpus are timeouts, and routing recurrence is intuitively well-suited to this specific failure mode (where policies loop and expert selections repeat). Have the authors evaluated whether this signal remains discriminative under qualitatively different failure types—such as early object drops or incorrect grasps that terminate well before reaching the inference cap? If such failures are absent from LIBERO by construction, a detailed discussion of this boundary condition would help readers better assess how broadly this monitoring approach transfers beyond the benchmark. If my issue can be resolved in a reasonably satisfactory manner, I will consider increasing the score.

4 comments

Reply 1 - W1/Q1: Compact representations, transfer, and process dynamics

Thank you for recognizing the value of routing-based monitoring and paired intervention evaluation. We clarify three distinctions from the contemporaneous VLA-FAIL [1].

1. Compact computational readouts. Each routing site in our MoE policy provides probabilities over 32 experts and top-4 selections. NAR summarizes routing dynamics over layers, denoising steps, and execution history into a 54-dimensional descriptor. LLMD uses 1,024-dimensional last-layer features per action token and estimates a full covariance matrix for each token position (VLA-FAIL, §3.1, Tables 13-14). By comparing compact dynamics descriptors with a successful-reference bank, NAR avoids this per-token high-dimensional covariance estimation and inversion. It reuses routing records produced during action generation and requires no additional policy forward passes. The CPU implementation reported in Appendix C takes 3.14 ms per update on average.

2. One source suite supplies a transferable reference and threshold. We construct the reference bank and calibrate the threshold using only LIBERO-Goal successes, including native and perturbed Goal executions. In our retrospective cross-suite evaluation, keeping both fixed yields 97.2% failure recall at a 1.32% false-alarm rate across 4,320 native episodes from Object, Spatial, and long-horizon LIBERO-10, without target-suite reference fitting or threshold calibration (Table 2b). The full VLA-FAIL detector reports recall of 73.7%-99.0% for X-VLA and 77.5%-94.1% for π0.5 across its four LIBERO-Plus suites, with 100 successful rollouts used to calibrate thresholds separately for each suite (their Tables 11 and 17). These results distinguish the reference and calibration protocols; the policies and evaluation data differ between studies. Our results demonstrate reuse of a source-suite monitor without target-suite successes for refitting or recalibration.

3. Explicit relations within the computational process. LLMD measures deviations of current last-layer features from a reference distribution; ACC measures consistency between overlapping predictions in consecutive action chunks. NAR uses expert selections, routing probabilities, and participation statistics to characterize both denoising dynamics within action generation and changes and recurrence across policy inferences, comparing these descriptors with successful-execution references (§3, Appendix A). Pointwise deviation scores do not uniquely specify an execution history: two feature sequences can have identical Mahalanobis distances at every step but different recurrence across steps. This provides a principled motivation for explicitly modeling computational dynamics. Routing supplies a direct interface to expert selection and computational allocation, allowing these process relations to be represented compactly for online monitoring.

References

[1] Seligmann et al. (2026). VLA-FAIL: Efficient Task Failure Detection for Finetuned Vision-Language-Action Models. arXiv:2606.21386v2, §3.1, Tables 11, 13–14, 17.

Reply 2 - W2/Q2: Failure processes and architectural scope

1. Timeout is the terminal label for unfinished tasks, not a single physical failure mode. Under our evaluator, intermediate events such as unsuccessful grasps and object drops do not immediately terminate an episode; execution continues. NAR monitors routing dynamics without requiring predefined failure types. We annotated physical events in 300 saved LIBERO-10 episodes (200 failures and 100 successes). The original NAR alarmed on 193/200 failed episodes; Table R1 groups those failures by observed execution process.

Table R1. Original NAR alarm coverage among failed episodes containing each event.

Physical event or operational proxyAlarmed / failed episodesCoverage
Empty closure (no observed contact at sampled boundaries)11/11100.0%
No lift after near-object closure124/12896.9%
Sustained immobility82/82100.0%
Low realized motion despite nonzero translation commands82/82100.0%

Categories may overlap. Coverage indicates whether NAR alarmed at any point in an eventually failed episode containing that event; it is not event-specific classification accuracy. The shared timeout label thus encompasses distinct physical execution processes.

2. Routing also supports finer-grained diagnosis of physical events. We trained supervised linear probes on current and past routing at the first policy query following event confirmation, using configuration-grouped five-fold cross-validation. Table R2 reports discrimination of drops after grasp, empty closures, and failed lifts.

Table R2. Routing-based event diagnosis under configuration-grouped five-fold CV.

Physical eventEvent / control samplesAUROCCorrectly identified events
Drop after established grasp7 / 1501.0006 / 7
Empty closure11 / 1500.99910 / 11
No lift after near-object closure133 / 1500.991129 / 133

The event counts are based on a fixed 0.5 decision threshold; positive and control samples are defined by local physical events, not final episode outcomes. Together, these results indicate that routing traces contain event-related information beyond binary execution alarms, motivating finer-grained diagnosis and targeted recovery.

3. Architectural scope and extension. Our experiments use HiMoE-VLA. The broader idea of monitoring execution through internal dynamics could extend to dense models through neuron activations. NAD (Neuron Agreement Decoding) selects LLM responses using activation sparsity and cross-sample neuron agreement, while RAD (Routing Agreement Decoding) selects candidates using agreement in MoE routing [4,5]. Together, these studies show that both neuron activations and expert routing can provide useful internal decision signals. They motivate exploring whether analogous signals in dense and MoE embodied policies can support execution monitoring and recovery.

The development of MoE language models and embodied architectures such as HiMoE-VLA and AdaMoE [2,3] also motivates our view of sparse computation as an important direction for scaling embodied models. Understanding and monitoring MoE policies therefore has both immediate research relevance and value for future embodied models. Our intervention-evaluation framework based on rescue, harm, and break-even precision is itself independent of discrete routing and can be applied to other monitoring signals (§4).

References

[2] Du et al. (2025). HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies. arXiv:2512.05693.

[3] Shen et al. (2025). Expertise Need Not Monopolize: Action-Specialized Mixture of Experts for Vision-Language-Action Learning. arXiv:2510.14300.

[4] Chen et al. (2026). Do LLMs Signal When They're Right? Evidence from Neuron Agreement. ICML, PMLR 306.

[5] Chen et al. (2026). Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control. arXiv:2606.22798.

Thank you for the thorough and well-evidenced responses to both questions. The new physical event annotations and probe results in Reply 2 are particularly convincing. Based on the rebuttal, I am raising my score from 2 to 3.

Thank you for carefully considering our responses and updating your assessment. We are glad that you found the additional physical event annotations and probe results convincing, and we appreciate your constructive questions and thoughtful feedback.

Reviewer_xC6hOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes NAR, a test-time harness for detecting execution failures and guiding recovery in frozen VLA models with an MoE architecture. NAR leverages the recurrence of expert routing history generated during model inference as an internal signal to detect execution stalls or failures. Experiments in the LIBERO environment demonstrate that routing information can detect failures with higher recall than monitoring external physical states or actions.

Critical strengths
  • The approach of using MoE for monitoring is highly practical: it does not require additional sensor information or model forward passes, while keeping computational costs low.
Critical weaknesses
  • All experiments are conducted in simulation environments like LIBERO and CALVIN, lacking validation on actual robot hardware, which makes it difficult to understand in which cases the method can be applied and in which cases it cannot.
  • The mechanism for dynamically selecting optimal recovery actions is unsolved.
  • The paper presents a solution applicable to the MoE architecture of VLA models, which is not common so far.
Questions for authors
  • Is the insight of the paper applicable to dense models, which are more common than current VLA models?
  • Is MoE routing information sufficient to detect which kind of anomaly is happening, other than the occurrence of the anomaly?
2 comments

Thank you for recognizing NAR's practical value. We clarify its generalization, recovery utility, and ability to reveal specific execution events.

W1: Simulation and real-robot relevance

Simulation enables reproducible evaluation across tasks, initial states, and perturbations. NAR generalizes to new LIBERO states, held-out tasks, and suites (Table 2), with evaluation data excluded from reference construction and calibration. SIMPLER also reports strong sim-to-real performance correlation when visual and control gaps are reduced [3].

EvaluationSize
LIBERO: new initial states2,560 episodes
LIBERO: held-out tasks3,200 episodes
LIBERO: held-out suite5,760 episodes
LIBERO-Plus perturbations25,902 episodes
CALVIN D→D548 subtasks

In the first three settings, recall is 86.9–95.3%, with 0.61–0.93% episode-level false alarms on successes. These results motivate testing NAR on physical robots. We are particularly interested in how routing signals behave under real contact, sensing noise, and complex interactions.

W2: Choosing recovery actions

Optimal recovery requires interpreting execution history and the current state, then predicting how alternative actions change future execution and success. Counterfactual outcomes, rescue–harm trade-offs, and limited budgets make this difficult. We have studied fixed-response utility, budget, and trigger settings (Tables 11–13), and explored routing-based selection over 13 primitives, 8 directions, and 256 candidates (Table 14).

With the pooled profile, the fixed release-and-lift action R(0.60,1) produces 175 net additional successes across 1,608 completed alarm pairs, a +10.88 percentage-point gain per completed alarm (95% CI [4.67,16.83]; Table 11), accounting for both rescue and harm against unchanged continuation. We are interested in using routing dynamics to better estimate candidate outcomes and select recovery actions.

W3 / Q1: Dense models and MoE relevance

Dense VLAs lack discrete expert routing, but analogous internal dynamics may be studied through neuron activations. In LLMs, NAD uses neuron-activation agreement and RAD uses MoE routing agreement to select responses [4,5], motivating analogous investigations in embodied policies. The development of MoE LLMs and MoE VLAs such as HiMoE-VLA and AdaMoE [1,2] also makes routing-based monitoring increasingly relevant.

Q2: Can routing identify anomaly types?

Yes. NAR alarms cover several physical failure patterns. Among 200 failed episodes in our 300-episode LIBERO-10 cohort, we measured episode-level alarm coverage for annotated events:

Table R1. Original NAR alarm coverage for annotated events in failed episodes.

Physical event or proxyAlarmed / failedCoverage
Empty closure (no sampled contact)11/11100.0%
No lift after near-object closure124/12896.9%
Sustained immobility82/82100.0%
Insufficient motion despite translation commands82/82100.0%

Coverage means an alarm occurred at any point in a failed episode containing that event; event classes may overlap.

For type identification, separate supervised binary probes read only routing available at the first policy query after event confirmation and earlier. In configuration-grouped five-fold CV, they correctly identify 6/7 held-object drops, 10/11 empty closures, and 129/133 no-lift events (AUROCs 1.000, 0.999, 0.991; 150 eligible negatives per type). Matching positives and negatives within tasks and nearby query times yields AUROCs 0.972, 1.000, 0.957 on 6, 10, and 83 pairs. Thus, routing carries information about concrete execution events, providing a basis for diagnosis-informed recovery.

References

[1] Du et al. (2025). HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies. arXiv:2512.05693.

[2] Shen et al. (2025). Expertise Need Not Monopolize: Action-Specialized Mixture of Experts for Vision-Language-Action Learning. arXiv:2510.14300.

[3] Li et al. (2025). Evaluating Real-World Robot Manipulation Policies in Simulation. CoRL, PMLR 270.

[4] Chen et al. (2026). Do LLMs Signal When They're Right? Evidence from Neuron Agreement. ICML, PMLR 306.

[5] Chen et al. (2026). Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control. arXiv:2606.22798.

Dear Reviewer xC6h,

Thank you again for your time and feedback. We would very much appreciate hearing whether there are any concerns that our response has not fully addressed. We would be glad to elaborate on our reasoning and discuss any remaining questions with you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.