acceptodds
Under review as a conference paper at ICLR 2027

TMT: Runtime Backdoor Detection for Vision-Language-Action Policies on Unseen Tasks

12 views

Abstract

Backdoored vision-language-action (VLA) policies can preserve benign task performance while producing malicious actions when a trigger appears. Detecting such activation is difficult because malicious behavior can comprise individually plausible actions, while unfamiliar tasks introduce legitimate changes in observations and behavior. We introduce TMT, a runtime backdoor detector based on Token Manifold and Latent Transition modeling. Trained on benign rollouts, its two branches assess input-token structure and prediction errors in adjacent-layer latent dynamics. A suspicious rollout identified by the token manifold branch, once confirmed through latent deviations, guides transition selection for subsequent monitoring. We further explore policy purification through self-distillation: a frozen copy of the backdoored policy provides benign-input actions to supervise a student on paired benign and triggered observations, without requiring a separate clean reference policy. For evaluation, we adapt traditional backdoor detectors and repurpose anomaly and failure detection methods as VLA backdoor detectors. In a post-hoc comparison with ten baselines, TMT achieves state-of-the-art backdoor detection performance on unseen tasks across three VLA backdoor attacks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-Language-Action Models, Backdoor Detection, Runtime Monitoring

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

This paper proposes TMT for runtime backdoor detection in vision-language-action policies, with an emphasis on unseen tasks. A token-manifold branch models benign input representations, while a latent-transition branch predicts changes between adjacent backbone layers. Either branch can trigger an alarm. Before subsequent monitoring, an incident-discovery procedure uses a suspicious rollout to select the monitored transition. The paper also explores policy purification through self-distillation on paired benign and triggered observations. Experiments cover three attacks and three victim-policy families across the reported settings, ten adapted baselines, and real-world robot execution. After incident-driven transition selection, the unseen-task evaluation on OpenVLA-OFT reports TDR/FRR of 100.00%/0.57% for BadVLA, 83.43%/8.57% for GoBA, and 73.60%/0.00% for DropVLA.

Critical strengths
  1. Important problem, practical design, and broad evaluation. VLA backdoors can cause unsafe physical behavior, making this a timely problem. Combining benign representation modeling with layer-transition prediction on a frozen policy is a reasonable, implementable design. The evaluation spans several attack and victim settings, ten baseline adaptations, robot experiments, and a purification extension. The baseline adaptation effort is substantial, and the robot study provides preliminary evidence of deployment feasibility.
  2. Candid and informative ablations. The authors explicitly report that the token-manifold branch reaches only 0.57% and 15.20% TDR on unseen GoBA and DropVLA tasks, respectively, and acknowledge that unseen-task detection mainly relies on latent transitions. This transparency is valuable and helps identify the actual source of the gains.
Critical weaknesses
  1. Incident-driven adaptation leaves the threat model and attack dependence unclear. Section 3 excludes triggered training examples, but Section 4.3 and Appendix F use a suspicious incident to select the monitored transition. Even if predictor fitting uses only benign data, transition selection remains part of detector adaptation. Not reading malicious labels is not equivalent to avoiding attack-related observations. The discovery data source, its relationship to the evaluated attack, and the fallback when no incident is accepted are insufficiently specified. It is also unclear whether selection is repeated for each attack/checkpoint; stability across incidents and seeds and transfer across attacks are not analyzed. If selection requires an incident from the same attack, the results establish cross-task detection after attack-related adaptation. This limits the scope of the central generalization claim.
  2. Branch complementarity is limited in the main unseen-task setting. On unseen GoBA tasks, adding the token-manifold branch leaves TDR unchanged at 83.43% and latency unchanged at 0.28 queries, while increasing FRR from 7.43% to 8.57%. On unseen DropVLA tasks, the combination likewise leaves TDR unchanged at 73.60%, although latency improves from 6.25 to 5.68 queries. These results support the latent-transition branch, but provide limited evidence for complementary detection coverage during unseen-task monitoring. The manifold branch may still help discover the initial incident; that role should be distinguished from its incremental value during subsequent monitoring.
Questions for authors
  1. Discovery and benign-only evaluation. Where do discovery rollouts come from, and what happens if no incident is accepted? What performance remains when fitting, selection, and calibration all exclude triggered observations?
  2. Stability and attack dependence. Is selection repeated for each attack/checkpoint? How stable is the selected transition across incidents and seeds, and can it transfer to another attack without reselection?
  3. Paper and code correspondence. Please clarify how the released SAFE-ID implementation maps to TMT, particularly the two-query window and discovery-score aggregation, and whether Equation (8) uses the raw or calibrated transition score.
1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the careful review and constructive feedback. We provide additional clarifications and experimental evidence below.


W1 & Q1: …incident-driven adaptation…discovery data source…what happens if no incident is accepted?

Data source. Both predictors are trained using only benign inference-time rollouts from the frozen, potentially backdoored VLA policy. The tasks corresponding to these rollout samples refer to the seen tasks for TMT.

We clarify that our unseen-task evaluation measures cross-task generalization after incident-driven transition selection and threshold calibration, rather than detection without transition selection.

No-incident setting. The token-manifold branch alone achieves 95.41–100% TDR on seen tasks (seen setting). For unseen tasks, latent-transition monitoring provides broader detection coverage. We therefore additionally explored benign-anomaly-based transition selection as a potential fallback without using triggered observations for selection.

Transition-selection ablation. We compared the original, second-ranked, and benign-anomaly-selected transitions.

Backdoored policySelectionFRR ↓TDR ↑
BadVLAOriginal0.57%100.00%
BadVLASecond-ranked0.86%100.00%
BadVLABenign anomaly1.14%100.00%
GoBAOriginal8.57%83.43%
GoBASecond-ranked10.86%67.43%
GoBABenign anomaly14.57%49.14%
DropVLAOriginal0.00%73.60%
DropVLASecond-ranked1.60%44.00%
DropVLABenign anomaly1.60%40.00%

Notably, even when a benign rollout is used for transition selection, TMT still achieves the highest TDR among all baselines in Table 1. We have included these additional results in the Appendix D.2.


W1 & Q2: …stability across incidents and seeds…transfer across attacks without reselection?

The results in Table 1 are averaged over three independent random seen/unseen-task splits. Across all three splits, TMT consistently selects the same transition for each attack: layer 30 for BadVLA, layer 17 for GoBA, and layer 20 for DropVLA. This demonstrates stable transition selection across different task splits.

Our additional ablation further shows that even when a transition is selected wrong, TMT retains competitive detection performance, achieving the highest or joint-highest TDR among the baselines in Table 1.

The three attacks use distinct triggers, and we observed 0% backdoor activation when applying one attack's trigger to another attack's backdoored policy. Accordingly, TMT independently models the benign token manifold and latent transitions of each potentially backdoored policy, with transition selection performed separately for each model. Cross-attack transfer without reselection is therefore not part of our policy-specific detection setting.


W2: …branch complementarity is limited in the main unseen-task setting…

We agree that the two branches play different roles during detection. The token-manifold branch provides strong detection on seen tasks and is responsible for discovering suspicious incidents, while the latent-transition branch provides the main detection coverage on unseen tasks.

However, in real-world deployment, robots may repeatedly execute previously seen tasks while also encountering unseen tasks. Therefore, we recommend keeping both branches active during runtime monitoring to maintain detection capability across both settings, rather than relying solely on the latent-transition branch. Their complementarity lies in supporting different task settings, rather than necessarily improving TDR within the unseen-task setting alone.


Q3: …SAFE-ID implementation…two-query window…discovery-score aggregation…Equation (8)…

Thank you for requesting this clarification. SAFE-ID was the original name of the method now presented as TMT. The two-query window is an implementation detail of online detection in the latent-transition branch, while discovery-score aggregation belongs to the incident-discovery procedure for transition selection.

We evaluated all compared methods within a shared codebase developed for VLA backdoor benchmarking. Due to the supplementary material size limit, we extracted the relevant detector and transition-selection modules for submission. We plan to release the complete benchmarking framework after standardizing the implementation details, hoping to facilitate future research on VLA backdoor attacks and defenses.

Equation (8). Section 4.4 introduces the calibrated transition score s~D\tilde{s}_D, whereas Equation (8) uses sDs_D. We have corrected this notation (line 274).


We hope our clarifications and additional experiments have adequately addressed your concerns, and we would be very grateful if you might consider raising your score in light of these updates.

Reviewer_Q4kKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies runtime backdoor detection for VLA policies on unseen tasks. The authors propose TMT, which combines token manifold and latent transition modeling to detect deviations from benign input-token structure and adjacent-layer representation dynamics. The detector learns from benign rollouts without requiring prior knowledge of the backdoor trigger patterns or a separate clean reference for the OOD detection. The paper also explores policy purification through self-distillation using paired benign and triggered observations. Experiments compare TMT with existing backdoor, anomaly, and failure detection methods under three VLA backdoor attacks.

Critical strengths

(1) The paper addresses an important challenge in VLA deployment: detecting trigger-induced backdoor behavior while accommodating legitimate execution on unseen tasks. Its attack settings include both execution disruption and goal redirection, where individual malicious actions may resemble benign actions.

(2) Monitoring internal token representations and layer-wise transitions is a potentially useful approach that leaves the victim policy unchanged. Predictor fitting uses benign rollouts without explicit trigger specifications. The branch ablations identify the latent transition branch as the main source of unseen-task detection coverage for GoBA and DropVLA.

Critical weaknesses

(1) The selectivity of the learned anomaly (or OOD) signals for backdoor activation remains insufficiently established. Although the unseen-task experiments demonstrate benign/triggered OOD samples detection in the evaluated settings, one dataset separation is not sufficient. Additionally, it does not isolate backdoor activation from trigger appearance or generic representation shifts. The ambiguity also affects incident-driven transition selection: a score-confirmed incident is not independently verified as malicious, and benign anomalies could potentially influence the selected monitoring layer. How it may affect the method's performance is not thoroughly evaluated. Matched controls and evaluation of benign incident candidates would strengthen the proposed mechanism.

(2) Robustness to broader benign distribution shifts is insufficiently characterized. The LIBERO-10 task split provides valid cross-task evidence, but does not systematically vary object, visual, or instruction novelty. The 8.57% FRR under unseen-task GoBA still motivates controlled shift experiments and analysis of false alarms.

Questions for authors

(1) What evidence distinguishes backdoor activation from trigger appearance or generic OOD activation in the learned scores? Could the authors provide matched policy/input controls beyond the current LIBERO-10 task split? In addition, can benign anomalous rollouts satisfy the incident-confirmation criteria, and how would selection based on such an incident affect subsequent detection?

(2) How do detection rates and false rejections change under controlled benign object, visual, and instruction shifts? Please analyze the benign rollouts responsible for the 8.57% FRR under unseen-task GoBA and clarify whether these errors are associated with particular tasks or conditions.

(3) How robust is TMT to attacks that explicitly align triggered representations with benign distributions, as motivated by distribution-preserving attacks in self-supervised learning (Distribution Preserving Backdoor Attack in Self-supervised Learning, Tao et al., IEEE S&P 2024)? Could the authors evaluate at least one adaptive VLA attack?

(4) Could the authors clarify the scope of the detection generalization claims and provide stronger evidence distinguishing backdoor activation from benign distribution shifts? Or an improved overall claim, such as "agnostic backdoor detection in VLA", together with a substantive response to these concerns, could lead me to revise my score upward.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the constructive feedback.


W1 & Q1: …trigger appearance or generic representation shifts…matched policy/input controls…

To distinguish backdoor activation from trigger appearance and generic shifts, we conducted two control experiments.

Input controls. We added a novel object ("fake trigger") to test scenes to distinguish trigger appearance from attack-associated alarms.

Backdoored policyFake-trigger SRFake-trigger FRRTrue-trigger TDR
GoBA89.14%4.57%83.43%
BadVLA92.86%0.57%100.00%
DropVLA90.40%2.40%73.60%

Fake-trigger FRRs are 0.57%–4.57%, versus 73.6%–100% true-trigger TDRs (Table 1), showing TMT does not indiscriminately flag these novel objects (OOD/representation shifts).

Policy controls. We fitted TMT on a clean policy under the same seen-unseen protocol, with and without the novel object.

Clean-policy conditionOriginal-task SRFRR
Seen, no novel object90.67%0.00%
Seen, novel object89.33%0.00%
Unseen, no novel object92.57%6.57%
Unseen, novel object84.86%9.42%

FRR was 0% on seen tasks and most unseen benign rollouts remained unflagged with the novel object. This supports TMT modeling benign token representations and transitions rather than novelty alone. Details are added to Appendix D.3 and D.4.


W1 & Q1: …benign anomalous rollouts…incident-confirmation criteria…affect subsequent detection?

Transition-selection ablation. We compared the original, second-ranked, and benign-anomaly-selected transitions (This unlikely in real-world deployment, as suspicious rollouts flagged by TMT should undergo manual review to ensure safety.).

Backdoored policySelectionFRR ↓TDR ↑
BadVLAOriginal0.57%100.00%
BadVLASecond-ranked0.86%100.00%
BadVLABenign anomaly1.14%100.00%
GoBAOriginal8.57%83.43%
GoBASecond-ranked10.86%67.43%
GoBABenign anomaly14.57%49.14%
DropVLAOriginal0.00%73.60%
DropVLASecond-ranked1.60%44.00%
DropVLABenign anomaly1.60%40.00%

Benign-anomaly selection raises FRR by 0.57, 1.60, and 6.00 pp for BadVLA, DropVLA, and GoBA, respectively. However, TMT still achieves the highest TDR among all baselines in Table 1. Details are added to Appendix D.2.


W2 & Q2: …benign object, visual, and instruction shifts…8.57% FRR…

LIBERO-10. Beyond LIBERO-10, we evaluated TMT and the baselines on real-world robot tasks. We chose LIBERO-10 for simulation because it combines diverse manipulation challenges found across the LIBERO benchmark, including different objects, scene layouts, and multi-step goals. These factors make LIBERO-10 particularly suitable for evaluating generalization to substantially different unseen tasks.

False-rejection analysis. The 8.57% GoBA FRR averages three random task splits (7.71%, 8.29%, 9.71%). We further examined the false-rejected benign rollouts and found that they were distributed across most unseen tasks, without an obvious concentration in any particular task. These false rejections occurred in both successful and unsuccessful benign rollouts, with no clear task-specific failure pattern observed.


Q3: …align triggered representations with benign distributions…adaptive VLA attack?

We agree this is a relevant threat. NAD [1] and ARGD [2] motivate a possible trade-off between benign alignment and attack effectiveness. Single-layer alignment need not remove downstream deviations; all-layer alignment might erase the attack. We have not found an open-source VLA implementation of this kind of attack, and we have added a discussion in Limitations.

[1] Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks, ICLR 2021.

[2] Eliminating Backdoor Triggers for Deep Neural Networks Using Attention Relation Graph Distillation, IJCAI 2022.


Q4: …scope of the detection generalization claims…

We thank the reviewer for this suggestion. Our work focuses on distinguishing benign execution from backdoor activation on unseen tasks, while the detector is trained only on benign rollouts from seen tasks. This setting does not require prior knowledge of trigger patterns or a separate clean reference policy. Furthermore, our additional policy- and input-control experiments demonstrate that TMT does not indiscriminately flag the novel inputs. We have further clarified in the newly added sections that our findings apply to the evaluated attacks and distribution shifts, rather than guaranteeing robustness against untested adaptive attacks.


Together, we hope our detailed responses and the additional experiments have adequately addressed your concerns, and we would be very grateful if you might consider raising your score in light of these updates.

While the authors did solid work of more ablation studies that does improve the claim, my main concerns are still unanswered convincingly. I still reckon overall score of the paper is below the ICLR bar.

Thank you for your response. Could you please let us know which concerns remain unconvincing? Since we still have some time, we hope to do our best to address your concerns and improve the quality of our work.

Reviewer_BWiYOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors introduce TMT, a runtime backdoor detector for VLA policies, that fits on benign rollouts from a frozen victim policy. The way TMT works is that two branches run per policy query/input: a token manifold branch scores input tokens with flow matching and a latent transition branch scores error prediction on adjacent layer hidden state differences. There is an alarm that fires when either branch exceeds its calibrated threshold. The authors evaluate TMT across various experimental settings. Overall I like the paper. I have some minor concerns and comments that I list below.

Critical strengths
  • I really like the problem and how it is framed with the four requirements, especially 2) that malicious behavior can consist of individually plausible actions distinguished only by which goal the sequence serves.

  • The unseen task generalization is included in the paper and it greatly strengthens the robustness of TMT.

  • I like the wide variety of experimental settings and how the baselines are categorized into backdoor detection and anomaly and failure detection. Overall, the paper is easy to read, the writing is clear and figures are comprehensive.

Critical weaknesses
  • To me it seems that TMT requires white box access to the model because it says 'can inspect the policy's internal representations during inference'. It should be clearly written if the setting is white box or blackbox.

  • I dont see a proper limitations section. After 6.5 the paper goes straight to the conclusion. At the minimum, the paper should talk about the white-box requirement, adaptive attacks etc.

  • Some experiments to show adaptive attacker would be good. Specifically, if the adaptive attacker adds a penalty keeping triggered token representations near the benign manifold and triggered layer transitions near benign dynamics. This is because the premise of the detector is that triggers induce representational deviation.

Questions for authors
  • Is there any part of TMT that would be compatible with a fully blackbox setting?

  • Does detection survive an attacker that regularizes the triggered representations toward the benign manifold?

  • Does purification require the trigger, and if so, how is it obtained in deployment?

1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

We sincerely thank the reviewer for their careful review and encouraging feedback. We appreciate your recognition of our problem formulation and unseen-task evaluation, and clarify the access requirements and adaptive-attack considerations below.


W1, W2 & Q1: …requires white box access…compatible with a fully black-box setting? …a proper limitations section…

TMT requires access to the internal representations of the potentially backdoored VLA policy and is therefore not a fully black-box detector. specifically, this access is read-only during detection: the victim policy remains frozen, and TMT does not require its original training process or an uncompromised reference policy.

We have added a new section before the conclusion to discuss the existing limitations and explicitly highlight the requirement to access the internal representation.


W3 & Q2: …keeping triggered token representations near …Does detection survive…?

We agree that representation-aligned adaptive attacks are a relevant threat. The findings of NAD [1] and ARGD [2] motivate our hypothesis that there may be a trade-off between the degree of benign alignment and the effectiveness of a representation-aligned adaptive attack.

Our latent transition branch learns benign adjacent-layer transitions across the backbone. Layer selection occurs after the attacker has produced the backdoored model. Therefore, the attacker may not know in advance which transition will exhibit the largest deviation and be selected. If the monitored transition deviates sufficiently from benign dynamics, TMT can raise an alarm. Conversely, if the representation-aligned attacks are aligned with benign inference across all layers, they may inhibit backdoor activation.

However, we have not identified an open-source available VLA attack implementation specifically targeting this joint-alignment objective. We explicitly discuss this untested adaptive threat in the newly added Limitations section (line 477-493).

[1] Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks, ICLR 2021.

[2] Eliminating Backdoor Triggers for Deep Neural Networks Using Attention Relation Graph Distillation, IJCAI 2022.


Q3: Does purification require the trigger, and if so, how is it obtained in deployment?

Yes, but purification performed after TMT flags a suspicious scenario; detection itself does not require prior knowledge of the trigger. However, in real-world deployment, TMT does not automatically identify or recover the trigger object, so human intervention is required to construct the paired benign–triggered environments. We have added this to Appendix H (line 1051-1053).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear AC and Reviewers,

We introduce TMT for runtime VLA backdoor detection on unseen tasks. Experiments show that TMT achieves SOTA performance against ten baselines across three VLA attacks. We thank the reviewers for recognizing the following strengths of TMT:

  • Practical method design (BWiY: S1; sZdG: S1; Q4kK: S2)
  • Unseen-task detection (BWiY: S2; Q4kK: S2; sZdG: S2)
  • Broad evaluation and solid ablations (BWiY: S3; sZdG: S1, S2; Q4kK: post-rebuttal comment)

We summarize the additional experiments and clarifications addressing the reviewers’ concerns below.


1. Incident dependence and transition-selection stability (sZdG: W1, Q1, Q2; Q4kK: W1, Q1)

We added two transition-selection experiments in Appendix D.2: selecting the second-ranked transition and selecting a transition using a benign rollout. Although TDR decreases for GoBA and DropVLA, both alternative selection strategies still achieve higher TDRs than all baselines in Table 1 for these two attacks, at their reported operating points. These experiments assess transition-selection stability and show that the detection advantage persists under these stress tests.

However, such erroneous selection is unlikely in practice, not only because our TMT has a low FRR of 0.6%–1.6% (Table 4), but also because, for safety reasons, a human should verify whether the robot is exhibiting malicious behavior whenever TMT raises an alert.


2. Distinguishing backdoor-associated alarms from benign novelty (Q4kK: W1, W2, Q1, Q2, Q4)

We added two sets of control experiments in Appendices D.3 and D.4. First, we introduced novel objects into scenes with backdoored policies. This creates benign novelty without activating the backdoor: observed FRRs are 0.57%–4.57%, compared with 73.6%–100% TDR when the actual triggers are present.

Second, we repeated the object controls using a clean policy with no backdoor. Observed FRRs are 6.57%–9.42% on unseen tasks and 0% on seen tasks. No false alarms were observed on seen tasks, including on unsuccessful benign rollouts.

These experiments show that benign novelty and task failures do not automatically trigger TMT alarms, strengthening the evidence beyond the original seen/unseen task comparison.


3. Representation-aligned adaptive attacks (BWiY: W3, Q2; Q4kK: Q3)

We have not identified an open-source VLA implementation targeting this joint-alignment objective. TMT learns adjacent-layer dynamics across the backbone. Consequently, alignment at one layer need not eliminate detectable deviations elsewhere. Prior work shows that aligning intermediate representations with a benign reference can suppress backdoors [1, 2], motivating a possible trade-off between representation alignment and attack effectiveness. Developing an adaptive VLA attack targeting this trade-off is beyond the scope of this paper. We discuss this untested threat in the new Limitations section.


4. Evaluation beyond LIBERO-10 (Q4kK: W2, Q2)

Beyond LIBERO-10, we evaluated TMT and the baselines on real-world robot tasks, extending the evaluation from simulation to physical deployment. We chose LIBERO-10 for simulation because it combines diverse manipulation challenges found across the LIBERO benchmark, including different objects, scene layouts, and multi-step goals. These factors are intertwined across challenging long-horizon tasks, making LIBERO-10 particularly suitable for evaluating generalization to substantially different unseen tasks.


5. Additional clarifications

  • Access requirements (BWiY: W1, W2, Q1). We explicitly state that TMT requires read-only internal access to the frozen policy and added a dedicated Limitations section.
  • Branch roles (sZdG: W2). We clarified that the manifold branch supports incident discovery and familiar-task detection, while the transition branch provides the main unseen-task coverage.
  • Implementation and notation (sZdG: Q3). We clarified SAFE-ID/TMT naming and the distinction between online two-query monitoring and incident discovery, and corrected Equation (8) to use the calibrated transition score.
  • Purification requirements (BWiY: Q3). We clarified in Appendix H that purification requires human-constructed pairs of benign and triggered observations. Detection requires no prior trigger specification or separate clean reference policy.

We respectfully invite the reviewers to reconsider their assessments in light of these additional experiments and clarifications. We welcome any remaining questions and thank you again for your time and thoughtful feedback.

Best regards,

The Authors

References

[1] Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks. ICLR, 2021.

[2] Eliminating Backdoor Triggers for Deep Neural Networks Using Attention Relation Graph Distillation. IJCAI, 2022.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.