acceptodds
Under review as a conference paper at ICLR 2027

SAFER: Selective FFN Readout for Cross-Task VLA Failure Detection

Abstract

Reliable failure detection is essential for deploying vision-language-action models (VLAs) across diverse manipulation tasks. Existing methods have explored using VLA hidden states to detect execution failures, but these aggregated representations do not reveal which internal computational contributions are most useful for failure detection on unseen tasks. We propose SAFER, a selective feed-forward network (FFN) readout method for cross-task VLA failure detection. The method uses successful and failed rollouts from the training set and performs training-free FFN neuron selection. It ranks neurons based on the separability of their activations between successful and failed rollouts and selects a subset of neurons for failure monitoring. At each execution step, we construct the detector input representation by accumulating the activation-weighted output directions of the selected neurons. The selected neuron subset is directly applied to unseen tasks, while the VLA parameters remain frozen and the original action computation process is preserved. Experiments show that SAFER can effectively improve failure detection performance on unseen tasks compared with full hidden-state readout.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-Language-Action Model, Failure Detection

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_LEy2OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies runtime failure detection for VLAs. The key idea is to replace the full hidden-state representation used by prior work with a selective readout from FFN neurons. To select the readabou, the authors compute a rollout-level activation statistic for each FFN neuron, rank neurons according to the absolute deviation, and reconstruct their activation-weighted output contribution through projection. This selective representation is then provided to an MLP or LSTM failure detector. The VLA itself remains frozen. The appraoch is empirically validated on LIBERO10 using four VLA policies.

Critical strengths
  • The proposed method is simple and computationally efficient. Neuron selection requires only rollout-level outcome labels from an offline dataset. It is substantially simpler than approaches requiring policy fine-tuning.

  • The empirical study is broad. The evaluation is across both autoregressive and flow-matching VLA architectures. The ablation study is also comprehensive and interesting. Reversing the ranking substantially degrades performance. The neuron-budget experiment is also informative, where using selected neurons outperforms reading all FFN neurons on unseen tasks

Critical weaknesses
  • Now the representation part needs two steps: first neuron selection, and then representation reconstruction. The experiments do not yet establish whether the gain comes more from representation construction or simply from supervised neuron selection. This is directly relevant to the main claim of this paper, so it is an important missing piece.

  • The neuron selector has access to the future consequence of failure, although the claimed application is online failure detection. SAFER selects neurons using the activation averaged over the entire rollout. But at deployment time, the detector is supposed to operate from a prefix to give timely detection/early warning, which is also the paper's emphasis.

Questions for authors
  • What happens if the detector receives the selected FFN activations (at,SK(a_{t,S_K} directly rather than the reconstructed one with FFN? Is the output-direction reconstruction actually important?

  • How different are the results if neuron selectivity is computed within each task and then aggregated across training tasks? This seems important for ruling out task-identity confounding.

  • Are SAFE and SAFER hyperparameters tuned over identical search spaces and identical budgets? If so, please specify the hyperparameter search spaces.

5 comments
AuthorsOpenReview ↗

Response to Reviewer LEy2 (1/3)

[W1][Q1]. [W1]Now the representation part needs two steps: first neuron selection, and then representation reconstruction. The experiments do not yet establish whether the gain comes more from representation construction or simply from supervised neuron selection. This is directly relevant to the main claim of this paper, so it is an important missing piece. [Q1]What happens if the detector receives the selected FFN activations directly rather than the reconstructed one with FFN? Is the output-direction reconstruction actually important?

A: We thank the reviewer for raising this important question. Table R1 shows that both supervised neuron selection and FFN output-direction reconstruction contribute to SAFER's performance gains. Neuron selection provides discriminative information, while output-direction reconstruction further improves detection performance.

We compare SAFER-Raw, which directly uses AUC-selected neuron activations, with the full SAFER using reconstructed representations. With the same selected Top-100 neurons, reconstruction further increases the average Unseen AUC to 80.74% for MLP and 67.79% for LSTM.

These results indicate that neuron selection provides performance gains, while representation reconstruction using the pretrained FFN output projection weights offers additional improvements. The updated results are reported in Table 6 of the revised paper.

Table R1. Effect of Neuron Selection and FFN Reconstruction.

\hline \scriptscriptstyle\text{Method} &\scriptscriptstyle\text{OpenVLA-Seen} &\scriptscriptstyle\text{OpenVLA-Unseen} &\scriptscriptstyle\pi_0\text{-Seen} &\scriptscriptstyle\pi_0\text{-Unseen} &\scriptscriptstyle\text{WidowX-Seen} &\scriptscriptstyle\text{WidowX-Unseen} &\scriptscriptstyle\text{Avg.-Seen} &\scriptscriptstyle\text{Avg.-Unseen}\\\\ \hline \scriptscriptstyle\text{SAFE-LSTM} &\scriptscriptstyle\mathbf{63.34}&\scriptscriptstyle\mathbf{65.33}&\scriptscriptstyle59.46&\scriptscriptstyle63.91&\scriptscriptstyle81.14&\scriptscriptstyle69.31&\scriptscriptstyle67.98&\scriptscriptstyle66.18\\\\ \scriptscriptstyle\text{Raw Random-100-LSTM} &\scriptscriptstyle54.20&\scriptscriptstyle54.91&\scriptscriptstyle70.32&\scriptscriptstyle66.30&\scriptscriptstyle48.39&\scriptscriptstyle46.96&\scriptscriptstyle57.64&\scriptscriptstyle56.08\\\\ \scriptscriptstyle\text{SAFER-Raw-LSTM} &\scriptscriptstyle58.91&\scriptscriptstyle61.04&\scriptscriptstyle73.23&\scriptscriptstyle65.80&\scriptscriptstyle75.26&\scriptscriptstyle65.25&\scriptscriptstyle69.13&\scriptscriptstyle64.03\\\\ \scriptscriptstyle\text{SAFER-LSTM} &\scriptscriptstyle61.89&\scriptscriptstyle60.94&\scriptscriptstyle\mathbf{76.64}&\scriptscriptstyle\mathbf{69.84}&\scriptscriptstyle\mathbf{85.31}&\scriptscriptstyle\mathbf{72.60}&\scriptscriptstyle\mathbf{74.61}&\scriptscriptstyle\mathbf{67.79}\\\\ \hline \scriptscriptstyle\text{SAFE-MLP} &\scriptscriptstyle\mathbf{68.78}&\scriptscriptstyle62.93&\scriptscriptstyle70.00&\scriptscriptstyle75.12&\scriptscriptstyle88.69&\scriptscriptstyle87.78&\scriptscriptstyle75.82&\scriptscriptstyle75.28\\\\ \scriptscriptstyle\text{Raw Random-100-MLP} &\scriptscriptstyle64.26&\scriptscriptstyle57.44&\scriptscriptstyle45.48&\scriptscriptstyle51.44&\scriptscriptstyle79.96&\scriptscriptstyle77.82&\scriptscriptstyle63.23&\scriptscriptstyle62.23\\\\ \scriptscriptstyle\text{SAFER-Raw-MLP} &\scriptscriptstyle68.28&\scriptscriptstyle71.05&\scriptscriptstyle77.08&\scriptscriptstyle77.00&\scriptscriptstyle89.16&\scriptscriptstyle\mathbf{89.19}&\scriptscriptstyle78.17&\scriptscriptstyle79.08\\\\ \scriptscriptstyle\text{SAFER-MLP} &\scriptscriptstyle68.32&\scriptscriptstyle\mathbf{71.60}&\scriptscriptstyle\mathbf{77.83}&\scriptscriptstyle\mathbf{81.43}&\scriptscriptstyle\mathbf{89.44}&\scriptscriptstyle89.18&\scriptscriptstyle\mathbf{78.53}&\scriptscriptstyle\mathbf{80.74}\\\\ \hline \end{array}$$
AuthorsOpenReview ↗

Response to Reviewer LEy2 (2/3)

[W2] The neuron selector has access to the future consequence of failure, although the claimed application is online failure detection. SAFER selects neurons using the activation averaged over the entire rollout. But at deployment time, the detector is supposed to operate from a prefix to give timely detection/early warning, which is also the paper's emphasis.

A: Thank you for pointing out this issue. To examine the effect of selecting neurons using activations averaged over full rollouts, we conducted additional experiments with different trajectory prefix lengths. In Table R2, we evaluate SAFER on OpenVLA and π₀ using prefix lengths ranging from 200 steps to the full rollout. The average successful rollout lengths are 284.5 and 260.5 steps for OpenVLA and π₀, respectively.

At Prefix-300, SAFER already achieves meaningful cross-task failure detection performance. However, full-rollout neuron selection still achieves the highest average Unseen AUC. These results suggest that shorter trajectory prefixes already contain useful signals for failure detection. In contrast, full rollouts provide more information for offline neuron selection and further improve cross-task detection performance. These findings support the feasibility of our neuron selection strategy. The updated results are reported in Table 9 of the revised paper.

Table R2. Effect of Trajectory Prefix Length on Downstream Failure Detection Performance.

\hline \scriptstyle\text{Method} &\scriptstyle\text{OpenVLA-Seen} &\scriptstyle\text{OpenVLA-Unseen} &\scriptstyle\pi_0\text{-Seen} &\scriptstyle\pi_0\text{-Unseen} &\scriptstyle\text{Avg.-Seen} &\scriptstyle\text{Avg.-Unseen}\\\\ \hline \scriptstyle\text{SAFE-MLP} &\scriptstyle68.78&\scriptstyle62.93&\scriptstyle70.00&\scriptstyle75.12&\scriptstyle69.39&\scriptstyle69.03\\\\ \scriptstyle\text{SAFER-MLP (Prefix-200)} &\scriptstyle68.16&\scriptstyle69.66&\scriptstyle72.29&\scriptstyle76.83&\scriptstyle70.23&\scriptstyle73.25\\\\ \scriptstyle\text{SAFER-MLP (Prefix-250)} &\scriptstyle72.14&\scriptstyle71.00&\scriptstyle68.51&\scriptstyle72.41&\scriptstyle70.33&\scriptstyle71.71\\\\ \scriptstyle\text{SAFER-MLP (Prefix-300)} &\scriptstyle\mathbf{74.92}&\scriptstyle68.13&\scriptstyle75.08&\scriptstyle79.91&\scriptstyle\mathbf{75.00}&\scriptstyle74.02\\\\ \scriptstyle\text{SAFER-MLP} &\scriptstyle68.32&\scriptstyle\mathbf{71.60}&\scriptstyle\mathbf{77.83}&\scriptstyle\mathbf{81.43}&\scriptstyle73.08&\scriptstyle\mathbf{76.52}\\\\ \hline \scriptstyle\text{SAFE-LSTM} &\scriptstyle\mathbf{63.34}&\scriptstyle65.33&\scriptstyle59.46&\scriptstyle63.91&\scriptstyle61.40&\scriptstyle64.62\\\\ \scriptstyle\text{SAFER-LSTM (Prefix-200)} &\scriptstyle60.63&\scriptstyle58.29&\scriptstyle72.52&\scriptstyle70.77&\scriptstyle66.58&\scriptstyle64.53\\\\ \scriptstyle\text{SAFER-LSTM (Prefix-250)} &\scriptstyle62.60&\scriptstyle49.48&\scriptstyle75.29&\scriptstyle68.04&\scriptstyle68.95&\scriptstyle58.76\\\\ \scriptstyle\text{SAFER-LSTM (Prefix-300)} &\scriptstyle59.72&\scriptstyle57.63&\scriptstyle75.36&\scriptstyle64.73&\scriptstyle67.54&\scriptstyle61.18\\\\ \scriptstyle\text{SAFER-LSTM} &\scriptstyle61.89&\scriptstyle\mathbf{60.94}&\scriptstyle\mathbf{76.64}&\scriptstyle69.84&\scriptstyle\mathbf{69.27}&\scriptstyle\mathbf{65.39}\\\\ \hline \end{array}$$
AuthorsOpenReview ↗

Response to Reviewer LEy2 (3/3)

[Q2] How different are the results if neuron selectivity is computed within each task and then aggregated across training tasks? This seems important for ruling out task-identity confounding.

A: We thank the reviewer for this valuable suggestion. We recomputed neuron rankings using Within-Task AUC on OpenVLA-LIBERO10 and π₀-LIBERO10, and compared the selected neurons with those from the original Pooled-AUC selection. Table R3 shows that the average Top-100 overlap is 50.67% for OpenVLA and 42.33% for π₀. These results indicate that the two selection methods identify partially overlapping sets of neurons. We have added these results to Table 13 of the revised appendix.

Table R3. Top-100 Neuron Overlap Between Pooled-AUC and Within-Task AUC Selection.

\hline \scriptstyle\text{Benchmark}&\scriptstyle\text{Top-100 Overlap (\\%)}\\\\ \hline \scriptstyle\text{OpenVLA-LIBERO10}&\scriptstyle50.67\\\\ \scriptstyle\pi_0\text{-LIBERO10}&\scriptstyle42.33\\\\ \hline \end{array}$$ We further compared SAFER's detection performance under the original Pooled-AUC and Within-Task AUC selection methods. **Table R4 shows that Within-Task AUC selection does not substantially reduce SAFER's detection performance.** The original Pooled-AUC selection performs better with LSTM detectors, while MLP detectors achieve comparable results under both selection methods. These results suggest that SAFER's detection capability does not rely solely on differences in failure rates across tasks. We have added these results to Table 12 of the revised appendix. **Table R4.** Within-Task AUC Neuron Selection Ablation. $$\begin{array}{lcccccc} \hline \scriptstyle\text{Method} &\scriptstyle\text{OpenVLA-Seen} &\scriptstyle\text{OpenVLA-Unseen} &\scriptstyle\pi_0\text{-Seen} &\scriptstyle\pi_0\text{-Unseen} &\scriptstyle\text{Avg.-Seen} &\scriptstyle\text{Avg.-Unseen}\\\\ \hline \scriptstyle\text{SAFER-MLP (Within-Task)} &\scriptstyle70.91&\scriptstyle72.79&\scriptstyle73.46&\scriptstyle81.48&\scriptstyle72.19&\scriptstyle\mathbf{77.14}\\\\ \scriptstyle\text{SAFER-MLP (Original)} &\scriptstyle68.32&\scriptstyle71.60&\scriptstyle77.83&\scriptstyle81.43&\scriptstyle\mathbf{73.08}&\scriptstyle76.52\\\\ \hline \scriptstyle\text{SAFER-LSTM (Within-Task)} &\scriptstyle58.82&\scriptstyle58.60&\scriptstyle74.83&\scriptstyle65.98&\scriptstyle66.83&\scriptstyle62.29\\\\ \scriptstyle\text{SAFER-LSTM (Original)} &\scriptstyle61.89&\scriptstyle60.94&\scriptstyle76.64&\scriptstyle69.84&\scriptstyle\mathbf{69.27}&\scriptstyle\mathbf{65.39}\\\\ \hline \end{array}$$ >**[Q3]** Are SAFE and SAFER hyperparameters tuned over identical search spaces and identical budgets? If so, please specify the hyperparameter search spaces. **A:** Sorry for the confusion. We use the same hyperparameter tuning protocol for SAFE and SAFER, with identical search spaces for learning rates and regularization coefficients in matched comparisons. We provide the complete hyperparameter search spaces in the appendix to ensure reproducibility. We have added these details to Table 36 and 37 of the revised appendix.
Reviewer_84ZoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SAFER, a selective FFN readout method for cross-task failure detection in vision-language-action models. Rather than feeding the final-layer features directly into a failure detector as SAFE, SAFER extracts selective FFN representations and uses them as inputs to the classifiers adopted in SAFE, which are then trained for failure detection. The results show that SAFER improves failure detection performance in many settings, although performance decreases in some cases.

Critical strengths

Demonstrates improved performance over SAFE in certain cases. Introduces FFNs to the field of failure detection and shows the utility of using FFN-related terms from internal model computations for failure monitoring.

Critical weaknesses
  1. Novelty and significance of the contribution

If the main contribution is simply to replace the input representations of the same classifier used in SAFE with FFN-derived features, while retaining the underlying premise of detecting failures from internal representations, the proposed method appears to offer an incremental improvement rather than a substantial contribution. As currently presented, the contribution seems to amount primarily to feeding FFN-derived representations into the SAFE detector. Are there additional contributions or deeper scientific insights beyond this engineering modification that I may have overlooked?

  1. Inconsistent performance and strength of the scientific claims

SAFER-LSTM appears to exhibit performance degradation in several cases, and the boldface formatting in Table 1 seems to obscure this trend. In particular, MolmoAct2 exhibits an approximately 20-percentage-point drop in performance. These results suggest that SAFER provides an alternative engineering choice that can improve performance in certain settings, rather than establishing a generally effective approach to failure detection. I am therefore not convinced that the current results support a sufficiently strong scientific claim for a venue such as ICLR.

  1. Robustness of neuron selection

In the "Selector Stability Across Split Seeds" experiment, the reported overlap rates of approximately 20–30% are higher than the random baseline, but still seem low in absolute terms. This raises concerns about the robustness and reproducibility of the neuron selection procedure. Further evidence is needed to establish that the selected neurons are sufficiently stable across different training splits and that the method's performance does not depend excessively on the particular data split used for selection. Furthermore, given the low overlap between selected neuron subsets across different data splits, the authors report results based on a single experimental run. Since the method does not require full-scale LLM training, the computational cost of repeating the experiments is likely to be manageable. I therefore encourage the authors to evaluate the method across multiple independent runs and report the results to demonstrate the robustness and reproducibility of their findings.

Questions for authors
  1. Fundamental assumption underlying the approach

I have a question about the fundamental assumption underlying this work. Why should the outputs of Transformer blocks provide useful signals for failure monitoring in VLA models? From an information bottleneck perspective, information irrelevant to task execution should be discarded from the learned representations. I am therefore uncertain whether the proposed approach reflects a general principle that can be expected to hold broadly, or whether its effectiveness is a heuristic observation that happens to hold for the particular neural network architectures and policies evaluated in this paper.

  1. Clarification of the training-free claim

If I understand the method correctly, although the abstract emphasizes training-free FFN neuron selection, the readout detector itself still needs to be trained, as illustrated in Fig. 2, similarly to SAFE. If so, the abstract seems somewhat overstated, as it may give readers the impression that the overall failure detection method is training-free. I believe the paper should more clearly distinguish between training-free neuron selection and the training required to fit the failure detector.

  1. Fairness of the comparison with SAFE

Does LIBERO10 refer to a different dataset from LIBERO, on which SAFE was evaluated? If so, I would like to know how SAFE performs on the LIBERO benchmark used in the original evaluation. Such results would help establish a fairer and more direct comparison between SAFER and SAFE.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 84Zo (1/4)

[W1,Q1] W1:Novelty and significance of the contribution. If the main contribution is simply to replace the input representations of the same classifier used in SAFE with FFN-derived features, while retaining the underlying premise of detecting failures from internal representations, the proposed method appears to offer an incremental improvement rather than a substantial contribution. As currently presented, the contribution seems to amount primarily to feeding FFN-derived representations into the SAFE detector. Are there additional contributions or deeper scientific insights beyond this engineering modification that I may have overlooked?[Q1]Fundamental assumption underlying the approach: I have a question about the fundamental assumption underlying this work. Why should the outputs of Transformer blocks provide useful signals for failure monitoring in VLA models? From an information bottleneck perspective, information irrelevant to task execution should be discarded from the learned representations. I am therefore uncertain whether the proposed approach reflects a general principle that can be expected to hold broadly, or whether its effectiveness is a heuristic observation that happens to hold for the particular neural network architectures and policies evaluated in this paper.

A: We thank the reviewer for raising this fundamental question. Although VLA models are trained for action prediction rather than explicit failure detection, their internal representations contain information about task progress, spatial relationships between the robot arm and objects, and action-related states. This information may also provide useful signals for failure detection. SAFE (Gu et al., SAFE, NeurIPS 2025. ) has already demonstrated the feasibility of detecting failures from VLA internal representations. However, its appendix visualizations show that the shared "Failure Zone" observed in the main paper does not consistently emerge across different policies and tasks.

Why FFN Neurons? Recent mechanistic interpretability research by Häon et al. (Häon et al., Mechanistic Interpretability for Steering Vision-Language-Action Models, CoRL 2025. ) showed that FFN neurons in OpenVLA and π₀ encode interpretable control-related concepts and can causally influence robot behavior.Inspired by this finding, SAFER investigates whether failure-discriminative information is concentrated in specific FFN neurons. We use AUC to select neurons with strong failure-discriminative capabilities and reconstruct their output contributions using the pretrained FFN output projection weights (WdownW_{\mathrm{down}}). We retain the same MLP/LSTM detectors as SAFE.

Our key finding is that selective FFN representations reveal more consistent cross-task failure-related structures across different VLA policies. In Figure 5 (π₀-LIBERO10), SAFER substantially reduces task-dependent clustering compared with SAFE. Representations from the later stages of failed rollouts become more concentrated across tasks, forming a clearer shared "Failure Zone“. Additional qualitative analysis results for other VLA policies are provided in Appendix A.10 of the revised manuscript. This suggests that the phenomenon is not limited to a single VLA policy.

Furthermore, our quantitative experiments demonstrate SAFER's advantages in cross-task generalization to Unseen tasks. These findings provide new empirical evidence about how failure-discriminative information is organized at the neuron level in VLA models. Therefore, SAFER's contribution goes beyond replacing the input features of an existing detector. It also reveals the value of selective FFN readout for cross-task failure monitoring.

AuthorsOpenReview ↗

Response to Reviewer 84Zo (2/4)

[W2] Inconsistent performance and strength of the scientific claims SAFER-LSTM appears to exhibit performance degradation in several cases, and the boldface formatting in Table 1 seems to obscure this trend. In particular, MolmoAct2 exhibits an approximately 20-percentage-point drop in performance. These results suggest that SAFER provides an alternative engineering choice that can improve performance in certain settings, rather than establishing a generally effective approach to failure detection. I am therefore not convinced that the current results support a sufficiently strong scientific claim for a venue such as ICLR.

A: We thank the reviewer for pointing out this issue. SAFER-LSTM shows substantial performance variation on MolmoAct2 Seen tasks. This is likely related to the small validation sets and the very limited number of failed rollouts. Table R1 shows that each of the three task-level random splits contains 50 Seen validation rollouts. Among them, only 2, 2, and 1 rollouts are failures, respectively. Under such severe class imbalance, ROC-AUC is highly sensitive to the prediction rankings of a few failed rollouts. This can lead to large performance variations across seeds.

However, Table R2 shows that SAFER-LSTM achieves a higher average Unseen AUC than SAFE-LSTM (60.39% vs. 50.68%), although it does not improve performance on every seed. This indicates that the performance degradation observed on Seen tasks does not consistently extend to Unseen tasks. In addition, SAFER shows more favorable results with MLP detectors. We have added these details to Table 22 and Table 23 of the revised appendix.

Table R1. Seed-Level SAFE-LSTM vs. SAFER-LSTM Performance on MolmoAct2 Seen Tasks.

\hline \scriptstyle\text{Seed} &\scriptstyle\text{Seen S/F} &\scriptstyle\text{SAFE-LSTM} &\scriptstyle\text{SAFER-LSTM} &\scriptstyle\Delta\\\\ \hline \scriptstyle0&\scriptstyle48/2&\scriptstyle62.50&\scriptstyle87.50&\scriptstyle+25.00\\\\ \scriptstyle1&\scriptstyle48/2&\scriptstyle86.46&\scriptstyle62.50&\scriptstyle-23.96\\\\ \scriptstyle2&\scriptstyle49/1&\scriptstyle100.00&\scriptstyle40.82&\scriptstyle-59.18\\\\ \hline \scriptstyle\text{Mean}& &\scriptstyle82.99&\scriptstyle63.61&\scriptstyle-19.38\\\\ \hline \end{array}$$ **Table R2.** Seed-Level SAFE-LSTM vs. SAFER-LSTM Performance on MolmoAct2 Unseen Tasks. $$\begin{array}{ccccc} \hline \scriptstyle\text{Seed} &\scriptstyle\text{Unseen S/F} &\scriptstyle\text{SAFE-LSTM} &\scriptstyle\text{SAFER-LSTM} &\scriptstyle\Delta\\\\ \hline \scriptstyle0&\scriptstyle42/8&\scriptstyle50.30&\scriptstyle64.88&\scriptstyle+14.58\\\\ \scriptstyle1&\scriptstyle29/21&\scriptstyle33.00&\scriptstyle61.08&\scriptstyle+28.08\\\\ \scriptstyle2&\scriptstyle48/2&\scriptstyle68.75&\scriptstyle55.21&\scriptstyle-13.54\\\\ \hline \scriptstyle\text{Mean}&-&\scriptstyle50.68&\scriptstyle60.39&\scriptstyle+9.70\\\\ \hline \end{array}$$
AuthorsOpenReview ↗

Response to Reviewer 84Zo (3/4)

[W3] Robustness of neuron selection: In the "Selector Stability Across Split Seeds" experiment, the reported overlap rates of approximately 20–30% are higher than the random baseline, but still seem low in absolute terms. This raises concerns about the robustness and reproducibility of the neuron selection procedure. Further evidence is needed to establish that the selected neurons are sufficiently stable across different training splits and that the method's performance does not depend excessively on the particular data split used for selection. Furthermore, given the low overlap between selected neuron subsets across different data splits, the authors report results based on a single experimental run. Since the method does not require full-scale LLM training, the computational cost of repeating the experiments is likely to be manageable. I therefore encourage the authors to evaluate the method across multiple independent runs and report the results to demonstrate the robustness and reproducibility of their findings.

A: Thanks to the reviewer for raising this concern. We would like to clarify that our results are not based on a single experimental run or random data splits within the same set of tasks. Our three split seeds use different random task-level partitions. We independently perform neuron selection, detector training, and evaluation for each seed (Seed 0/1/2). SAFER selects neurons using only the training data of each seed.

Table R3 shows that each seed contains only 210 training rollouts from 7 different tasks in LIBERO-10. Given the limited training data, the 20%–30% overlap among the selected Top-100 neurons is not perfect, but remains substantially higher than the random baseline among more than 10,000 FFN neurons. We have added these details to Table 20 and 21 of the revised appendix.

Furthermore, we report the cross-seed failure detection results of SAFE and SAFER across four VLA policies in Appendix A7. All Seen and Unseen results are reported as Mean ± Sample Std over three split seeds. Therefore, our main experimental results already account for variations across different task-level data splits and are not based on a single run.

Table R3. Per-Task Success and Failure Rollout Counts Across Three Seeds on OpenVLA- and π₀-LIBERO10.

\begin{array}{llccccccccc} \hline &&&\sz\text{Seed 0}&&&\sz\text{Seed 1}&&&\sz\text{Seed 2}&\\\\ \sz\text{Model}&\sz\text{Task}&\sz\text{Train}&\sz\text{Seen}&\sz\text{Unseen}&\sz\text{Train}&\sz\text{Seen}&\sz\text{Unseen}&\sz\text{Train}&\sz\text{Seen}&\sz\text{Unseen}\\\\ \hline \sz\text{OpenVLA}&\sz0&\sz-&\sz-&\sz27/23&\sz16/14&\sz11/9&\sz-&\sz16/14&\sz11/9&\sz-\\\\ &\sz1&\sz25/5&\sz14/6&\sz-&\sz22/8&\sz17/3&\sz-&\sz24/6&\sz15/5&\sz-\\\\ &\sz2&\sz20/10&\sz11/9&\sz-&\sz17/13&\sz14/6&\sz-&\sz19/11&\sz12/8&\sz-\\\\ &\sz3&\sz-&\sz-&\sz28/22&\sz19/11&\sz9/11&\sz-&\sz16/14&\sz12/8&\sz-\\\\ &\sz4&\sz8/22&\sz8/12&\sz-&\sz8/22&\sz8/12&\sz-&\sz10/20&\sz6/14&\sz-\\\\ &\sz5&\sz-&\sz-&\sz43/7&\sz-&\sz-&\sz43/7&\sz25/5&\sz18/2&\sz-\\\\ &\sz6&\sz13/17&\sz13/7&\sz-&\sz19/11&\sz7/13&\sz-&\sz-&\sz-&\sz26/24\\\\ &\sz7&\sz22/8&\sz8/12&\sz-&\sz-&\sz-&\sz30/20&\sz16/14&\sz14/6&\sz-\\\\ &\sz8&\sz5/25&\sz6/14&\sz-&\sz-&\sz-&\sz11/39&\sz-&\sz-&\sz11/39\\\\ &\sz9&\sz14/16&\sz10/10&\sz-&\sz18/12&\sz6/14&\sz-&\sz-&\sz-&\sz24/26\\\\ \hline \sz\pi_0&\sz0&\sz-&\sz-&\sz40/10&\sz27/3&\sz13/7&\sz-&\sz24/6&\sz16/4&\sz-\\\\ &\sz1&\sz28/2&\sz19/1&\sz-&\sz28/2&\sz19/1&\sz-&\sz28/2&\sz19/1&\sz-\\\\ &\sz2&\sz25/5&\sz15/5&\sz-&\sz24/6&\sz16/4&\sz-&\sz23/7&\sz17/3&\sz-\\\\ &\sz3&\sz-&\sz-&\sz50/0&\sz30/0&\sz20/0&\sz-&\sz30/0&\sz20/0&\sz-\\\\ &\sz4&\sz26/4&\sz17/3&\sz-&\sz24/6&\sz19/1&\sz-&\sz25/5&\sz18/2&\sz-\\\\ &\sz5&\sz-&\sz-&\sz49/1&\sz-&\sz-&\sz49/1&\sz30/0&\sz19/1&\sz-\\\\ &\sz6&\sz21/9&\sz18/2&\sz-&\sz22/8&\sz17/3&\sz-&\sz-&\sz-&\sz39/11\\\\ &\sz7&\sz28/2&\sz20/0&\sz-&\sz-&\sz-&\sz48/2&\sz30/0&\sz18/2&\sz-\\\\ &\sz8&\sz10/20&\sz1/19&\sz-&\sz-&\sz-&\sz11/39&\sz-&\sz-&\sz11/39\\\\ &\sz9&\sz23/7&\sz15/5&\sz-&\sz24/6&\sz14/6&\sz-&\sz-&\sz-&\sz38/12\\\\ \hline \end{array}$$ *Note: each cell denotes the number of successful/failed rollouts.*
Reviewer_56tEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SAFER, a runtime failure detector for VLA policies that changes what the detector reads from the frozen policy. SAFE feeds the detector the full hidden state. SAFER instead ranks the neurons of one FFN layer by how well their rollout-averaged activations separate successful from failed training rollouts. At each step it feeds the detector the sum of their contributions (each activation times its column of W_down). The selected neurons and the detector are then fixed and applied to unseen tasks. The detectors are SAFE's MLP and LSTM, so my understanding is that the main change from SAFE is the input representation. Experiments cover OpenVLA, pi0, pi0.5, and MolmoAct2 on LIBERO-10, plus SAFE's recorded OpenVLA-WidowX rollouts, with task-macro ROC-AUC as the main metric. The main result is that SAFER-MLP improves unseen-task AUC over SAFE-MLP by 9.2% on average (75.1 vs. 65.9). The paper also ablates K, the layer, the representation, the selection rule, the number of training tasks, and selection stability across seeds.

Critical strengths
  1. The method is simple and cheap. It only needs K columns of W_down and a small detector, it does not change the policy, and it should be easy to add to an existing SAFE-style pipeline.
  2. The evaluation covers four quite different policies (autoregressive and flow-matching) plus real-robot rollouts. The main comparison keeps the data, splits, and detector fixed, so it is easy to read.
  3. I like the move to task-macro AUC. Pooled AUC can reward differences in score level between tasks rather than detection within a task, and the pooled numbers are still given in the appendix.
Critical weaknesses
  1. I am not yet convinced that the gain comes from FFN contributions rather than from simply giving the detector a smaller, label-selected input, since mapping the selected activations through W_down is a fixed linear map that the detector's first layer could absorb, the full FFN write is worse than the full hidden state on unseen tasks (60.6 vs. 62.9 on OpenVLA and 62.7 vs. 75.1 on pi0), and on pi0 even 100 random neurons recover about half of the gain (62.7 to 72.3, vs. 81.6 with ranked selection).
  2. The selection rule seems a bit fragile to me, since it pools AUC across tasks, which can favor neurons that encode task identity when failure rates differ between tasks, and it averages activations over whole rollouts although failed rollouts are usually longer, so part of the ranking may reflect elapsed time or stalled behavior rather than failure itself.
Questions for authors
  1. Could the authors compare against simpler ways of shrinking the input? For example: the raw 100 selected activations, PCA or a random projection of the hidden state to about 100 dimensions, the same AUC ranking applied to hidden-state dimensions, and an L1 sparse probe, all with the same tuning budget and standardized features.
  2. How many failed and successful rollouts does each task have in each split?
  3. I have a small question about the selection statistic. How much longer are failed rollouts than successful ones, and how strongly do the selected neurons correlate with the time step? Does the ranking change if it only uses the first L_q steps, or a within-task AUC?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 56tE (1/5)

[W1] I am not yet convinced that the gain comes from FFN contributions rather than from simply giving the detector a smaller, label-selected input, since mapping the selected activations through W_down is a fixed linear map that the detector's first layer could absorb, the full FFN write is worse than the full hidden state on unseen tasks (60.6 vs. 62.9 on OpenVLA and 62.7 vs. 75.1 on pi0), and on pi0 even 100 random neurons recover about half of the gain (62.7 to 72.3, vs. 81.6 with ranked selection).

A: Thanks for raising this important question.

Ablation Study on FFN Output-Direction Reconstruction. In Table R1, we compare SAFER-Raw and SAFER using the same AUC-selected Top-100 neurons and detector configurations. SAFER achieves higher average Unseen AUC than SAFER-Raw. These results indicate that the selected FFN neuron activations provide discriminative information for failure detection. Furthermore, mapping these activations into their corresponding FFN output contribution space through the pretrained WdownW_{\mathrm{down}} provides additional performance gains in certain model and detector configurations. The updated results are reported in Table 6 of the revised paper.

Table R1. Effect of Neuron Selection and FFN Reconstruction.

\begin{array}{lcccccccc} \hline \z{Method}&\z{OpenVLA-Seen}&\z{OpenVLA-Unseen}&\q{\pi_0\text{-Seen}}&\q{\pi_0\text{-Unseen}}&\z{WidowX-Seen}&\z{WidowX-Unseen}&\z{Avg.-Seen}&\z{Avg.-Unseen}\\\\ \hline \z{SAFE-LSTM}&\w{63.34}&\w{65.33}&\q{59.46}&\q{63.91}&\q{81.14}&\q{69.31}&\q{67.98}&\q{66.18}\\\\ \z{Raw Random-100-LSTM}&\q{54.20}&\q{54.91}&\q{70.32}&\q{66.30}&\q{48.39}&\q{46.96}&\q{57.64}&\q{56.08}\\\\ \z{SAFER-Raw-LSTM}&\q{58.91}&\q{61.04}&\q{73.23}&\q{65.80}&\q{75.26}&\q{65.25}&\q{69.13}&\q{64.03}\\\\ \z{SAFER-LSTM}&\q{61.89}&\q{60.94}&\w{76.64}&\w{69.84}&\w{85.31}&\w{72.60}&\w{74.61}&\w{67.79}\\\\ \hline \z{SAFE-MLP}&\w{68.78}&\q{62.93}&\q{70.00}&\q{75.12}&\q{88.69}&\q{87.78}&\q{75.82}&\q{75.28}\\\\ \z{Raw Random-100-MLP}&\q{64.26}&\q{57.44}&\q{45.48}&\q{51.44}&\q{79.96}&\q{77.82}&\q{63.23}&\q{62.23}\\\\ \z{SAFER-Raw-MLP}&\q{68.28}&\q{71.05}&\q{77.08}&\q{77.00}&\q{89.16}&\w{89.19}&\q{78.17}&\q{79.08}\\\\ \z{SAFER-MLP}&\q{68.32}&\w{71.60}&\w{77.83}&\w{81.43}&\w{89.44}&\q{89.18}&\w{78.53}&\w{80.74}\\\\ \hline \end{array}$$ **Full FFN Write vs. Full Hidden State.** Full FFN Write applies $W_{\mathrm{down}}$ to all FFN neuron activations and produces a representation with the same dimensionality as Full Hidden State. However, these representations contain different information. Full FFN Write contains only the reconstructed FFN output, whereas Full Hidden State also incorporates contributions from the residual stream and attention branches. Normalization operations also participate in the Transformer computation. These differences help explain the weaker performance of Full FFN Write. In contrast, SAFER improves detection performance through selective neuron reconstruction. **Random Neuron Baseline on π₀.** Randomly selected FFN neurons already provide some useful signals for failure detection. This suggests that FFN activations may contain discriminative information even without supervised neuron ranking. However, the reconstructed random-neuron representation does not outperform SAFE-MLP on π₀, either on Unseen tasks (72.3% vs. 75.12%) or Seen tasks (58.90% vs. 70.00%). In comparison, AUC-based Top-100 neuron selection further improves the Unseen AUC to 81.6%. These results demonstrate the value of identifying FFN neurons with stronger failure-discriminative capabilities.
AuthorsOpenReview ↗

Response to Reviewer 56tE (2/5)

[Q1] Could the authors compare against simpler ways of shrinking the input? For example: the raw 100 selected activations, PCA or a random projection of the hidden state to about 100 dimensions, the same AUC ranking applied to hidden-state dimensions, and an L1 sparse probe, all with the same tuning budget and standardized features.

A: Thanks for raising this important question. Building on the experiments in Table 4 of our paper, we conducted additional comparisons with four hidden-state-based baselines: Hidden PCA-100 (reducing the full hidden state to 100 dimensions using PCA), Random Projection-100 (randomly projecting the full hidden state to 100 dimensions), Hidden Random Subset-100 (randomly selecting 100 dimensions from the hidden state), and Hidden AUC Top-100 (selecting the Top-100 hidden-state dimensions using the same AUC ranking criterion as SAFER). All methods use MLP detectors, standardized features, and the same hyperparameter tuning budget. Random Projection and Random Subset were each evaluated with 10 independent random repetitions.

Table R2 highlights the limitations of simple dimensionality reduction methods. In contrast, SAFER achieves the highest average Seen and Unseen ROC-AUC among all compared methods (78.53% and 80.74%, respectively). These results suggest that SAFER's performance gains cannot be attributed solely to dimensionality reduction. The updated results are reported in Table 20 of the revised appendix.

Table R2. Alternative Readout Representations and Feature-Selection Baselines.

\begin{array}{lcccccccc} \hline \z{Method}&\z{OpenVLA-Seen}&\z{OpenVLA-Unseen}&\q{\pi_0\text{-Seen}}&\q{\pi_0\text{-Unseen}}&\z{WidowX-Seen}&\z{WidowX-Unseen}&\z{Avg.-Seen}&\z{Avg.-Unseen}\\\\ \hline \z{SAFE (Full Hidden)}&\w{68.78}&\q{62.93}&\q{70.00}&\q{75.12}&\q{88.69}&\q{87.78}&\q{75.82}&\q{75.28}\\\\ \z{Random Projection-100}&\q{66.57}&\q{59.65}&\q{68.92}&\q{63.42}&\q{89.49}&\q{81.21}&\q{74.99}&\q{68.09}\\\\ \z{Hidden Random Subset-100}&\q{64.84}&\q{57.13}&\q{65.27}&\q{59.74}&\q{87.75}&\q{83.67}&\q{72.62}&\q{66.85}\\\\ \z{Hidden AUC Top-100}&\q{66.44}&\q{60.12}&\q{75.70}&\q{74.69}&\q{88.53}&\w{90.26}&\q{76.89}&\q{75.02}\\\\ \z{Hidden PCA-100}&\q{65.02}&\q{55.01}&\w{79.55}&\q{74.19}&\w{90.37}&\q{85.30}&\q{78.31}&\q{71.50}\\\\ \z{SAFER (Ours)}&\q{68.32}&\w{71.60}&\q{77.83}&\w{81.43}&\q{89.44}&\q{89.18}&\w{78.53}&\w{80.74}\\\\ \hline \end{array}$$
AuthorsOpenReview ↗

Title: Response to Reviewer 56tE (3/5)

[W2,Q3] [w2]The selection rule seems a bit fragile to me, since it pools AUC across tasks, which can favor neurons that encode task identity when failure rates differ between tasks, and it averages activations over whole rollouts although failed rollouts are usually longer, so part of the ranking may reflect elapsed time or stalled behavior rather than failure itself. [Q3]I have a small question about the selection statistic. How much longer are failed rollouts than successful ones, and how strongly do the selected neurons correlate with the time step? Does the ranking change if it only uses the first L_q steps, or a within-task AUC?

A: Thanks for suggesting a more comprehensive evaluation. We first examine the rollout length differences between successful and failed trajectories on OpenVLA-LIBERO10 and π₀-LIBERO10. We then evaluate the robustness of neuron selection and its effect on downstream failure detection from two aspects: trajectory prefix length (Prefix-(L_q)) and Within-Task AUC.

Rollout Length Statistics. We analyzed the rollout lengths of both VLA policies. The average successful rollout lengths are 284.5 steps for OpenVLA and 260.5 steps for π₀. All failed rollouts reach the maximum execution budget of 520 steps.

Effect of Prefix Length on Neuron Selection. We evaluated how the observation window affects neuron selection. In Table R3, we recomputed neuron rankings using only the first (L_q) steps of each rollout. The average Top-100 overlap increases with the prefix length on both benchmarks. At Prefix-300, the overlap reaches 44.67% for OpenVLA and 55.00% for π₀. These results indicate that neuron selection retains a degree of robustness to observation window truncation. The updated results are reported in Table 8 of the revised paper.

Table R3. Effect of Trajectory Prefix Length on Top-100 Neuron Selection Overlap.

\hline \scriptstyle\text{Prefix} &\scriptstyle\text{OpenVLA-LIBERO10} &\scriptstyle\pi_0\text{-LIBERO10}\\\\ &\scriptstyle\text{Overlap (\\%)} &\scriptstyle\text{Overlap (\\%)}\\\\ \hline \scriptstyle 200&\scriptstyle 20.67&\scriptstyle 27.67\\\\ \scriptstyle 250&\scriptstyle 29.67&\scriptstyle 42.00\\\\ \scriptstyle 300&\scriptstyle 44.67&\scriptstyle 55.00\\\\ \scriptstyle 520&\scriptstyle 100.00&\scriptstyle 100.00\\\\ \hline \end{array}$$ **Effect of Prefix Length on Detection.** We further evaluated MLP and LSTM detectors using neurons selected under different prefix lengths. Table R4 shows that SAFER already achieves useful cross-task failure detection performance at Prefix-300. However, the original SAFER using full rollouts still achieves the highest average Unseen AUC (76.52% for MLP and 65.39% for LSTM). These results suggest that SAFER's detection capability does not depend entirely on information from the later stages of failed rollouts. The updated results are reported in Table 9 of the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.