acceptodds
Under review as a conference paper at ICLR 2027

Natural State-Prediction Accuracy Can Hide Weak Controlled Responsiveness in VLA Readouts

Abstract

Accurately decoding object states from the internal representations of vision-language-action (VLA) models does not establish that the predictions respond faithfully to changes in the target physical state. In natural observations, object state, robot configuration, occlusion, and task progress vary together, allowing contextual cues to contribute to prediction. In this paper, we introduce an evaluation framework that separates prediction accuracy, target-state responsiveness, and context stability using physically validated observations that cross target coordinates with robot contexts. We demonstrate that high natural-trajectory accuracy can coexist with weak controlled target-state responsiveness in fixed representation–readout pairs. Comparisons and interventions involving representations, readouts, and training data show that the three properties provide distinct diagnostic information. Furthermore, adding responsiveness and context sensitivity to a failure predictor based on initial state error and physical variables reduces policy-failure prediction error on new initializations relative to the specified baseline while same-observation controlled MAE is also informative. These findings motivate evaluating target-state responsiveness and context stability alongside natural prediction accuracy, and examining their relationship to actual policy behavior and task outcomes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Robot Learning, Representation Learning, Physical Grounding, State Representation, Representation Evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wATBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper focuses on the problem of linear probes overfitting to variables which are simply correlated with the objective. For example one of the settings identified in the paper is the robot opening a drawer in a single scene. If you train a probe to output the position of the drawer you may expect that rather than learning to accurately reflect the state of the drawer you may overfit to the robot's position. The authors create a set of counterfactual data samples across two scenes to evaluate this across several VLAs and pretrained image encoders. Unsurprisingly, given the limited nature of the dataset the probes overfit to other visual aspects of the scene. Overall the main finding of this paper is well understood in probing and machine learning at large. Additionally the paper does not provide any fixes, nor test on any higher diversity scenes which may not suffer from this issue.

Critical strengths

This work is highly reproducible, the appendix is highly detailed and the code is included in the supplementary material.

Critical weaknesses

This paper is extremely difficult to understand and parse. This includes the presentation of the method and results in which the core contributions/findings of the paper are quite unclear. I feel this paper is not suitable for publication in its current state.

The experimental settings are quite limited in nature. Only applied to 2 simulated scenes. This is a major limitation of the paper. If the authors are trying to make general claims about the performance of linear probes a much much larger diversity of tasks is needed. Furthermore it is unclear if these models are adapted at all for these scenes. Nothing is reported on their natural success rates or if they are fine-tuned and this specific manipulator type.

It is confusing that the features/embeddings used to train the ridge regression and probes for task variables are trained only on pooled image activations from each of these models. Why should we expect that task relevant information lives in the image tokens rather than across the prompt or proprioception tokens ? For most VLAs, including the ones used in this paper the action expert attends to both image, prompt and proprioception tokens.

Questions for authors

Can you explain what checkpoints you use for the VLA models and if those models are able to complete the tasks that you collect your data from at a high success rate ?

Why do the authors only probe on pooled image tokens rather than other potentially more relevant action tokens ? Could this explain part of the findings of shortcut learning ?

3 comments
AuthorsOpenReview ↗

Response to Reviewer wATB

We thank the reviewer for the detailed comments and for recognizing the reproducibility of our work. Your feedback helped us clarify the empirical contribution, reorganize the presentation, and add analyses connecting state predictions to the policy's action-generation features and behavior.

1. Clearer contributions and presentation

The known risk that probes learn correlated cues motivates our study. We examine how predictions respond when measured object state and robot pose are changed separately in physically and visually validated observations, and how predictor training changes these responses.

The revised manuscript separates three properties: accuracy during task execution, response to state changes at a fixed robot pose, and stability across robot poses at a fixed state. The additional comparisons show why they must be evaluated together:

  • With the same 216 observations and labels, positive training weights, and shared normalization and regularization, changing state-pose weights changed prediction responses to robot pose.
  • With 72 observations and matched counts per state-pose combination, selecting state-varying rather than pose-varying observations improved state responsiveness but increased pose sensitivity in three allocation pairs. On the development panel, natural-MAE changes ranged from a decrease of 9.73 mm to an increase of 10.36 mm (Section 4.3; Appendices I.8, I.13, I.16-I.17).

These findings show how predictor training changes distinct prediction properties while VLA features remain fixed. The evaluation also includes nonlinear RBF and MLP predictors (Section 4.2; B.2).

We also clarified the practical implications of the existing training experiments. In one Faucet-SmolVLA development comparison, changing the allocation of 36 training observations reduced natural MAE from 0.5629 to 0.4693 rad (~17%) and controlled MAE from 0.3954 to 0.2648 rad (~33%). This provides a concrete example of improving both errors through training-data design (Section 4.5; C.2).

We rewrote the abstract, introduction, Figure 1, and methods around these distinctions. Sections 4.1-4.4 connect each finding to its evidence, Section 4.5 explains the observed improvements, and Section 5.2 states the scope of the conclusions.

2. Task coverage, checkpoints, and actual performance (Q1)

We now distinguish feature evaluation from autonomous policy evaluation.

  • Meta-World: We used frozen SmolVLA and OpenVLA, with π0 base and GR00T N1.6 in the model extension. Checkpoint identifiers are reported in Appendix A.6. We trained state predictors without adapting these policies to the Meta-World tasks. Natural trajectories came from scripted experts; autonomous success of these VLA checkpoints on those tasks was not established.
  • LIBERO Drawer: The policy is SmolVLA adapted for 40,000 steps. It reached the goal in 175/256 natural rollouts (68.4%) across 128 initial conditions.
  • LIBERO Plate: We used the frozen public pi05_libero checkpoint, described by its publisher as a 30,000-step LIBERO-finetuned π0.5 model. Five of 16 attempted development conditions met physical eligibility. All 10 natural and 20 controlled branches reached the goal, with branches sharing those five conditions and stopping at first goal attainment (A.8; F.1; I.2-I.3, I.9).

We distinguish Drawer/Faucet's central findings, Mug/Stove's boundary cases, and the added Plate analysis. Mug does not reproduce the accurate-natural/weak-response combination; Stove shows weak responsiveness with high success. Table 1 identifies each task's role. Broader validation across independent tasks remains future work (Section 5.2).

3. Features beyond pooled images (Q2)

We do not assume that image tokens contain all task-relevant information. We expanded the analysis to OpenVLA's final prompt features before the first action-token logits, and π0.5 prefix and action-expert features. Action-expert features are taken immediately before the output projection at the first vector-field call. We specify the actual inputs: the evaluated π0.5 prefix contains images and task text, without direct state or history tokens (Appendices I.3-I.4).

At one Plate development initial condition, six ridge/RBF predictors across visual, prefix, and action-expert features had predicted-to-actual state-change ratios from -0.015 to 0.303, where 1 is ideal. The action-expert ridge ratio was 0.166. The same frozen policy reached the goal in all four controlled and two natural branches from that condition.

These results show that weak state-readout responses are not confined to pooled image features in this setting. We report feature-specific responses alongside actual behavior, without treating a weak readout as evidence that the policy ignores physical state (Section 4.4; I.3-I.4).

Thank you for the response. I appreciate the additional features beyond pooled image tokens. However I do not feel that my primary concerns on diversity of experimental settings were fully addressed. I therefore maintain my original score.

AuthorsOpenReview ↗

Additional Microwave Pilot and Task Coverage

Thank you for clarifying the remaining concern. We agree that broader evaluation across tasks and environments is needed, as stated in Section 5.2. Our current contribution is the controlled evidence about readout properties and their training dependence within the evaluated settings.

To investigate your concern further, we completed an exploratory microwave-opening study using frozen π0.5-LIBERO features. All 40 controlled observations across eight initial conditions passed the specified physical and visual checks. For a 0.15-rad change in door angle, median absolute prediction changes were 0.0417 and 0.0644 rad in the two robot contexts. Changing robot context while holding the door angle fixed produced median prediction changes of 0.2797 and 0.3248 rad. Repeating the baseline produced identical predictions in all eight conditions.

Natural-development MAE was 0.0616 rad, compared with 0.6199 rad for the training-mean constant. Baseline error on the separate controlled cohort was substantially larger: 3.6070 rad versus 0.5625 rad for the constant. We report the attenuated state responses and context-dependent predictions together with this accuracy loss as preliminary evidence from an additional task.

The updated supplementary archive includes the complete protocol, all observation-level results, and reproducible numerical summaries in Additional_Task_Pilots.pdf and its accompanying files. Broader task validation remains an important direction for extending the scoped findings of this work.

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

This paper build a crossed observation protocol to study the problem that the VLA readouts may ignore observation change. They propose to fits state readouts to frozen VLA features, and then change the fixed robot context or object context. In their study of Meta-World drawer and faucet tasks,the readouts barely respond to state changes, where the controlled MAE is much larger than the natural MAE. The author also report that adding these diagnostics to a failure predictor slightly improves failure prediction on a LIBERO task.

Critical strengths
  • The diagnostic design is simple and easy to apply. With the crossed-observation protocol, it is straightforward to do a sanity check when trying to decode states from VLAs.
  • The empirical investigation in the experiments includes substantial controls on the observation with appropraite statistics and robustness analyses. Every controlled observation is physically and visually validated
Critical weaknesses
  • The failure in crossed observation may have different reasons, which may not be due to the VLA features themselves. This is not clearly studied in the paper. For example, the readoutmay just learn a shortcut from the robot pose and task progress, since it only sees the natural trajectories during training. As shown in G.4 of the appendix, a robot configuration baseline with no images can reach even lower natural MAE than VLA readouts. Also, this may be due to the out-of-distribution issue since such state–context combinations never appear in the natural training data, so low G can reflect a generic OOD failure rather than a lack of responsiveness.
  • The experimental setting is narrow and the behavioral relevance is weak. Specifically, the paper uses different environments, tasks and policies for representation analysis (using Meta-World) and behavioral analysis (using LIBERO). Thus, the paper never shows that the low responsiveness found in Meta-World affects any policy's behavior. In this case, both analyses should be run in the same setting. Also, the tasks used in the paper are too few, with only 1 LIEBRO task and 2 Meta-World tasks. Considering both Meta-World and LIBERO provides diverse tasks, more tasks should be evaluated to make the conclusion more convincing. Lastly, the failure-prediction gain is −0.010 on Brier on a single LIBERO task with 128 initializations. The proposed G and R also do not outperform plain controlled MAE (Table 5).
Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Response to Reviewer bBhD

We thank the reviewer for recognizing the simplicity of the protocol and the physical and visual validation. Your comments helped us clarify predictor training effects and connect state predictions with policy behavior.

1. Predictor training, contextual shortcuts, and distribution shift

We now explicitly define the object of evaluation as a frozen representation paired with a trained state predictor. The time and robot-configuration baselines show that task progress and robot variables can strongly predict state along natural trajectories. This motivates examining training correlations, but does not identify which cues the VLA policy or its predictor actually uses (Appendix G.4).

We added training comparisons while keeping the VLA features fixed. With matched training counts, replacing natural observations with controlled observations increased the state-response ratio from 0.038 to 0.758 for SmolVLA and from 0.036 to 0.862 for OpenVLA on the 32-condition Faucet panel, where 1 is ideal. Natural-observation error also increased. This shows that weak responsiveness depends on predictor training and is not an immutable property of the frozen features (I.8, I.13).

We then compared predictors fitted to the same 216 observations and labels, with positive weights for every observation, common normalization, and the same regularization. Changing only the state-pose weights changed how predictions differed between robot poses in both models (I.8, I.13).

A separate 72-observation comparison matched total and per-combination counts. On the development panel, three allocation pairs showed that selecting state-varying rather than pose-varying observations improved state responsiveness but increased pose sensitivity; natural-MAE changes depended on the allocation (I.16-I.17).

These comparisons establish effects of predictor training within frozen representations. We distinguish the shared observations in the weighting comparison from matched joint state-context support between natural training and controlled evaluation. The original natural-to-controlled gap is not fully separated into contextual shortcut and distribution-shift effects. Sections 4.3, 5.1, and 5.2 now explain what each comparison controls and supports.

2. Same-policy behavior, task coverage, and failure prediction

State prediction and behavior within the same policy. We clarified the existing LIBERO Drawer analysis: the same task-adapted SmolVLA supplies the natural prefix features, controlled diagnostics, and policy actions. The original panel contains 128 initial conditions and 256 rollouts, with 175 goal attainments (Table 1; Appendix I.9).

We also added a Plate analysis using the same frozen LIBERO-adapted π0.5 policy to extract visual, input-processing, and action-generation features and generate actions. At one development initial condition, six state predictors had state-response ratios from -0.015 to 0.303, while the same policy reached the goal in all four controlled and two natural branches. Across the broader development panel, five of 16 attempted conditions met the physical eligibility criteria, and all 30 branches reached the goal. Branches share initial conditions and stop at first goal attainment (Section 4.4; I.2-I.4).

These results show that weak predictor responsiveness can coexist with goal attainment within the same policy. We separately specify the Meta-World and LIBERO settings (Appendices A.8, I.9).

Task coverage. We distinguish the central findings and training comparisons in Drawer and Faucet, the boundary cases in Mug and Stove, and the added same-policy Plate analysis. Mug does not reproduce the accurate-natural/weak-response combination, while Stove shows weak responsiveness alongside high success. Additional Faucet comparisons strengthen evidence within that task. Table 1 and Sections 4.2 and 4.4 state each task's role; Section 5.2 discusses validation across more independent tasks.

Predictive value beyond controlled MAE. We completed an independent evaluation on 448 held-out initial conditions and 896 rollouts. Both prediction models used the same six controlled observations before rollout, with their fitted coefficients kept fixed. The baseline already included controlled MAE, initial prediction error, and physical variables.

Adding state-response error and context variation gave a Brier difference of -0.001635, augmented model minus baseline, with a 95% initial-condition bootstrap interval of [-0.006293, +0.003086]. Additional predictive value beyond controlled MAE was not established. We distinguish this from the original approximately -0.010 improvement over the initial-error and physical-variable baseline. Sections 4.4 and 5.2 and Appendices F.7 and I.1 report the comparison and its interpretation.

Reviewer_S7pUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper distinguishes natural decoding accuracy, target-state responsiveness, and context stability. Validated physical observations show that low natural-trajectory error can coexist with weak responsiveness. Further experiments compare models and readouts and examine relationships with policy failures.

Critical strengths
  1. Whether accurate decoding implies appropriate physical response is relevant to interpreting VLA representations.
  2. Physical and visual validation is substantial, covering multiple models and readouts.
  3. Statistical units are clear, and confirmatory, follow-up, and non-estimable results are carefully distinguished.
Critical weaknesses
  1. The source of weak responsiveness remains unresolved. Appendix G.1, Table S23 reports 0/32 eligible roots for the support-restricted natural-matched subsets in both Drawer and Faucet. Generalization failure on novel state–context combinations therefore remains possible. The authors acknowledge this, but controlled joint-support experiments are needed to distinguish contextual shortcuts from generalization effects.
  2. Additional predictive value over simple error is unestablished. Table S21 gives a 95% interval of [−0.010564,0.001869][−0.010564, 0.001869] for the MGR−MCM_{GR}−M_C Brier difference, crossing zero; MCM_C already uses MAE from the same controlled observations. The evidence supports informative additional observations, but does not establish why the proposed diagnostic summaries are needed for failure prediction.
  3. Task coverage of the central finding is limited. Table 4 reports Mug natural MAE of 128.314 mm, failing to reproduce accurate-natural/weak-response behavior. Appendix F.5 also shows weak Stove responsiveness alongside high policy success. These are useful boundary cases, but more independent tasks are needed to establish when the diagnostics are informative for control.
Questions for authors
  1. Does the finding persist with controlled joint training/evaluation support?
  2. Can a new independent evaluation establish predictive gains beyond controlled MAE?
  3. Under which task conditions does weak responsiveness relate to actual control failure?
1 comment
AuthorsOpenReview ↗

Response to Reviewer S7pU

We thank the reviewer for recognizing the relevance of our question, the physical and visual validation, and the clear definition of statistical units. Your comments helped us strengthen the experimental comparisons and clarify the interpretation of the results. We address each weakness together with its corresponding question below.

1. Training support and weak responsiveness (W1/Q1)

We kept the VLA features frozen and compared state predictors trained on natural and controlled observations with matched sample counts. On the 32-condition Faucet panel, the ratio of predicted to actual state change increased from 0.038 to 0.758 for SmolVLA and from 0.036 to 0.862 for OpenVLA, where 1 indicates an ideal response. Natural-observation error also increased, so we report accuracy and responsiveness together (Appendix I.13).

To isolate the effect of training weights, we fitted state predictors using the same 216 observations and labels, with positive weights for every observation and shared normalization and regularization. Changing only the state-pose weights changed how the predictors' outputs differed between robot poses in both models (I.8, I.13).

A separate comparison used 72 observations from 24 training initial conditions, with three observations per condition and equal counts for each nominal state-pose combination. Across three allocation pairs, selecting state-varying observations improved state responsiveness but increased pose sensitivity; the direction of the natural-MAE change depended on the allocation (I.16-I.17).

These comparisons show how predictor training affects responses while the VLA features remain fixed. We distinguish shared observations between the reweighting arms from matched joint state-context support between natural training and controlled evaluation. The latter was not established, and we retain the original 0/32 eligible initial conditions for the support-restricted natural-matched subsets in both tasks (Appendix G.1). Sections 4.3 and 5.2 explain these controls and the remaining distinction between contextual shortcuts and generalization effects.

2. Predictive value beyond controlled MAE (W2/Q2)

We completed a separate evaluation using 448 held-out initial conditions and 896 rollouts, with two rollouts per condition. Both failure prediction models used the same six controlled observations collected before rollout, and their fitted coefficients were kept fixed.

The baseline already included controlled MAE alongside initial prediction error and physical variables. Adding state-response error and context variation changed the mean Brier score from 0.195481 to 0.193846. The paired difference, augmented model minus baseline, was -0.001635, with a 95% interval of [-0.006293, +0.003086], obtained by resampling initial conditions while preserving the paired rollouts.

This evaluation did not establish additional predictive value beyond the baseline that already includes controlled MAE. We now distinguish this result from the original improvement over the baseline using initial error and physical variables. The diagnostic measures describe different prediction properties; their added value for failure prediction is evaluated separately. Sections 4.4 and 5.2 and Appendices F.7 and I.1 report the comparison and its interpretation.

3. Task coverage and relation to control failure (W3/Q3)

We clarified which findings each task supports. Mug remains a boundary case: its natural MAE of 128.314 mm does not reproduce the accurate-natural/weak-response combination. Stove shows that weak responsiveness can coexist with high policy success.

We added a Plate analysis linking state prediction and behavior within the same frozen policy, including visual, input-processing, and action-generation features. Five of 16 development conditions met the physical eligibility criteria, and all 30 branches from these conditions reached the task goal.

At one of these development initial conditions, six state predictors had predicted-to-actual state-change ratios from -0.015 to 0.303. The same policy reached the goal in all four controlled and two natural branches from that condition. All branches stopped at first goal attainment.

This provides direct evidence that weak predictor responsiveness alone does not imply control failure, including when predictors use action-generation features. We therefore report state prediction, controlled responsiveness, and actual goal attainment separately. These development results do not yet identify general task conditions under which weak responsiveness predicts failure. The task-specific evidence is reported in Sections 4.2 and 4.4 and Appendices E, F.5 and I.2-I.4; the scope of generalization is discussed in Section 5.2.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to All Reviewers: Revisions and New Findings

We thank all three reviewers for their careful reading and constructive comments. Their feedback motivated additional experiments and helped us clarify our main contributions.

1. Reviewers' main points

Reviewer S7pU asked why predictions respond weakly to actual state changes and whether additional measures improve failure prediction. Reviewer bBhD highlighted training data, unfamiliar evaluation conditions, links between state prediction and behavior within one model, and evaluation scope. Reviewer wATB emphasized readability, novelty, task competence, and features beyond the visual encoder.

2. Common issues and our approach

We examined how training data and predictor fitting affect state predictions, how predictions relate to the same policy's behavior, and whether diagnostic measures add value over strong baselines. A state predictor estimates an object's position or angle from frozen VLA features.

3. Experiments added after the initial submission

  • Training weights: Using the same 216 observations and labels, frozen VLA features, and other predictor settings, we changed only the weights of state-pose combinations (Appendices I.8, I.13).

  • Data selection: We compared selecting 72 observations by varying object state at a fixed robot pose or varying robot pose at a fixed object state within each initial condition. Total counts and counts per state-pose combination were matched across three allocation pairs (I.16-I.17).

  • Same-policy analysis: On Plate, we examined predictors using visual, input-processing, and action-generation features from the same frozen policy, together with goal attainment (I.2-I.4).

  • Independent failure prediction: Fixed failure predictors were evaluated on 448 held-out initial conditions, with two rollouts each (896 total). The baseline used mean absolute error (MAE) from the same controlled observations (I.1).

  • Action correction and selection: Correction methods used the same state estimates and data, with matched model size and training updates. Selectors were compared with the original policy, direct cost prediction, and action-only selection using goal attainment and control steps (I.23, I.31-I.33).

4. Manuscript revisions

We clarified the questions, comparison design, model adaptation, feature locations, and evaluation scope in the abstract, introduction, and methods. Sections 4.3-4.5 present the results; Section 5.2 and the appendices explain their scope and limitations.

Section 4.5 clarifies how existing experiments inform training data design. In one Faucet-SmolVLA development comparison with 36 training observations, changing state-pose pairings reduced MAE on task observations from 0.5629 to 0.4693 rad (~17%) and MAE on controlled observations from 0.3954 to 0.2648 rad (~33%). Both errors decreased and predictions followed state changes more closely in all eight allocation comparisons across SmolVLA and OpenVLA in this development evaluation (Appendix C.2). We also specify sample counts and conditions for mixing natural and controlled training observations (C.3).

The independent failure prediction evaluation did not establish an additional advantage from diagnostic measures over the MAE baseline. We report this result and its interpretation in Sections 4.4 and 5.2 and Appendices F.7 and I.1. Action-correction results are reported by image condition alongside the predefined evaluation criteria (I.23).

5. Additional findings and contributions

First, training weights changed prediction responses to robot pose even with identical data and frozen VLA features (Section 4.3; I.8, I.13).

Second, data selection affected accuracy, state responsiveness, and prediction stability across robot poses differently. In Drawer, selecting observations that varied object state improved average state responsiveness compared with varying robot pose, but also increased pose sensitivity at a fixed state. Across the three allocation pairs, the change in MAE on task observations ranged from a decrease of 9.73 mm to an increase of 10.36 mm. These three prediction properties therefore require separate evaluation (Section 4.3; I.16-I.17).

Third, weak state prediction responses coexisted with goal attainment by the same policy. For six predictors evaluated at one Plate development initial condition, predicted-to-actual state-change ratios ranged from -0.015 to 0.303, where 1 indicates an ideal response. Yet all six associated policy rollouts reached the goal. The comparison included predictors using action-generation features, supporting joint examination of prediction responses and actual behavior within the same policy (Section 4.4; I.3-I.4).

Individual responses provide details; complete results and conditions

AuthorsOpenReview ↗

Additional Task Pilot Following Reviewer wATB’s Suggestion

We agree with Reviewer wATB that broader task evaluation is important and appreciate the suggestion. We therefore conducted an additional exploratory microwave-opening study using frozen π0.5-LIBERO features.

All 40 controlled observations across eight initial conditions passed the specified physical and visual checks. For a 0.15-rad change in door angle, median absolute prediction changes were 0.0417 and 0.0644 rad in the two robot contexts. Changing robot context at fixed door angle produced larger median prediction changes of 0.2797 and 0.3248 rad.

These observations show attenuated state responses and context-dependent predictions in an additional task. The controlled cohort also exhibited substantial prediction error, which we report alongside the response measurements. We therefore present this as preliminary task-specific evidence, with broader validation remaining a future direction in Section 5.2.

The updated supplementary archive includes Additional_Task_Pilots.pdf, the complete observation-level results, and code reproducing the numerical summaries.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.