pith. machine review for the scientific record. sign in

arxiv: 2605.00321 · v1 · submitted 2026-05-01 · 💻 cs.RO

Recognition: unknown

Embodied Interpretability: Linking Causal Understanding to Generalization in Vision-Language-Action Models

Hanxin Zhang , Mingshuo Xu , Abdulqader Dhafer , Shigang Yue , Hongbiao Dong , Zhou Daniel Hao
Authors on Pith no claims yet

Pith reviewed 2026-05-09 19:41 UTC · model grok-4.3

classification 💻 cs.RO
keywords vision-language-actioncausal interpretabilitygeneralizationembodied policiesinterventional attributionnuisance featuresmanipulation tasks
0
0 comments X

The pith

Interventional attribution reveals when vision-language-action models depend on spurious features rather than true causes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Vision-language-action policies frequently fail when environments change because they base actions on irrelevant visual cues instead of the actual task elements. The work recasts the problem of attributing actions to visual inputs as one of estimating causal effects through interventions. It defines the Interventional Significance Score to quantify how much each visual region causally affects the predicted action and the Nuisance Mass Ratio to measure the fraction of influence coming from task-irrelevant parts. Theoretical analysis establishes that the masking method yields unbiased estimates and identifies when prediction error reliably indicates causal impact. Tests on multiple manipulation tasks confirm that the nuisance ratio forecasts generalization success while the significance score produces explanations that align better with actual causal structure than standard interpretability tools.

Core claim

We formulate visual-action attribution as an interventional estimation problem. Accordingly, we introduce the Interventional Significance Score (ISS), an interventional masking procedure for estimating the causal influence of visual regions on action predictions, and the Nuisance Mass Ratio (NMR), a scalar measure of attribution to task-irrelevant features. We analyze the statistical properties of ISS and show that it admits unbiased estimation, and we characterize conditions under which action prediction error provides a valid proxy for causal influence. Experiments across diverse manipulation tasks indicate that NMR predicts generalization behavior and that ISS yields more faithful explana

What carries the argument

The interventional masking procedure that computes the Interventional Significance Score (ISS) by measuring the effect on action predictions when specific visual regions are masked out, along with the Nuisance Mass Ratio (NMR) that aggregates attribution to irrelevant features.

Load-bearing premise

The interventional masking procedure yields unbiased estimates of causal influence on actions and action prediction error serves as a valid proxy for causal influence under the characterized conditions.

What would settle it

Running the experiments and finding no correlation between NMR values and actual generalization performance on new tasks, or finding that ISS explanations do not outperform baselines when compared to ground-truth causal interventions.

Figures

Figures reproduced from arXiv: 2605.00321 by Abdulqader Dhafer, Hanxin Zhang, Hongbiao Dong, Mingshuo Xu, Shigang Yue, Zhou Daniel Hao.

Figure 1
Figure 1. Figure 1: Generalization of VLAs can be analyzed through action attribution. For instance, in the task “stack the other cups on the top of the red cup”, failed trials rely more on nuisance visual cues (e.g., background, texture, and shadows) for decisions; successful trials rely more on task-relevant cues (e.g., manipulator, end-effector, and cups). when the visual inputs are completely masked (Omaisan & Mohamed, 20… view at source ↗
Figure 2
Figure 2. Figure 2: The overview of the proposed interpretability approach. Panel (A) illustrates the pipeline for generating the Interventional Significance Score (ISS), where action discrepancies (δk) induced by Bernoulli masking (Mk) and Gaussian mixing perturbations (Vblur) on multi-view observations are aggregated to yield saliency maps, subsequently forming a continuous ISS stream via linear interpolation. Panel (B) def… view at source ↗
Figure 3
Figure 3. Figure 3: Relationship between nmr@k and success rate. Each sample point corresponds to the success rate of a single task evalu￾ated under one random seed. For each of the 5 random seeds, 41 different tasks were evaluated, with each task executed over 25 trials to compute its success rate. (NORM), and provide detailed processing of these inter￾pretability baselines in Appendix E. 10−3 10−2 10−1 0.94 0.96 0.98 1 (0.0… view at source ↗
Figure 4
Figure 4. Figure 4: Optimal Robustness. ISS (red star) occupies the optimal top-right region, simultaneously maximizing the cosine similarity of the saliency map and minimizing the action’s MSE compared to other explanatory methods. Robustness study. We add Gaussian noise to the full-image input and select the top 5% episodes with the lowest action Mean Squared Error (MSE) across all tasks, yielding 200 episodes. Inspired by … view at source ↗
Figure 5
Figure 5. Figure 5: Saliency Fidelity Analysis. The bar chart displays the correlation between action MSE and saliency map changes across geometric, patch, and texture perturbations. ISS consistently shows a stronger linear alignment and higher Pearson coefficients than the ATT and NORM baselines. Fidelity study. We apply three types of structured pertur￾bations to nuisance regions across 1,000 randomly sampled episodes, incl… view at source ↗
Figure 6
Figure 6. Figure 6: ISS Visualization under perturbations. Saliency maps for the Close Jar task are visualized under four interventions (Gaussian noise, texture, geometric, and patch) at a fixed timestep. We report the variations of the nuisance mass ratio (nmr@10) under different perturbations. exhaustive intervention requires O(T × 3 × 162 ) forward passes, whereas our method reduces the cost to O(T × N). Consequently, whil… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of Episode Trajectories. Each plot represents the aggregated 3D spatial paths of the end-effector recorded across all episodes. The subfigures (a) through (d) illustrate four distinct manipulation tasks in the AGNOSTOS dataset view at source ↗
Figure 8
Figure 8. Figure 8: Segmentation masks for the “close jar” task (episode 0), showing 6 frames of front view. Green indicates action-relevant regions (Ωact), blue indicates task-support regions (Ωsup), and red indicates nuisance regions (Ωnuis). Fine-tuning. Implemented in the JAX framework, we fine-tuned the models using supervised fine-tuning (SFT) for 20 epochs on a single 96 GB NVIDIA RTX 6000 GPU with a batch size of 64. … view at source ↗
Figure 9
Figure 9. Figure 9: Optimization dynamics during VLA π0.5 training. C.2. Implementation and Evaluation. Implementation. We represent actions as continuous 8-DoF control signals. VLA π0.5 directly models continuous-valued actions via a flow-matching formulation, without discretizing the action space or requiring post-hoc decoding. The control frequency is set to 20 Hz (0.05 s per step), while policy inference is performed ever… view at source ↗
Figure 10
Figure 10. Figure 10: Correlation between nuisance mass ratio (nmr@k) and task success rate under different cutoff values k. The case k=10 is omitted here as it yields the strongest correlation and is presented in the main paper. under identical metric protocols. Scatter plots confirm that ISS maintains superior linearity across all splits. Under the Unseen geometric (GEO) intervention, ISS achieves a Pearson coefficient of r … view at source ↗
Figure 11
Figure 11. Figure 11: Saliency Fidelity Analysis between ∆ action and ∆ saliency across evaluation splits and perturbation types. Rows correspond to evaluation splits (Seen, Unseen, All), while columns correspond to different analysis settings. The first column summarizes Pearson correlation coefficients across perturbations. The remaining columns present scatter plots for ISS, ATT, and NORM under TEXTURE, GEO, and PATCH pertu… view at source ↗
Figure 12
Figure 12. Figure 12: Visualization of saliency maps and nuisance mass ratio ρ (k) ϕ (Ωnuis) with k = 10% under token magnitude (ϕ = | · |), attention score (ϕ = ATT), and interventional significance score (ϕ = ISS) for the “close jar” task across different camera views (Front, Wrist, Overhead). all attention heads to generate a unified attention map A¯ ∈ R N×N . Since the visual tokens in the backbone of VLA π0.5 utilize pref… view at source ↗
Figure 13
Figure 13. Figure 13: Representative episodes across seen and unseen tasks. 22 view at source ↗
read the original abstract

Vision-Language-Action (VLA) policies often fail under distribution shift, suggesting that decisions may depend on spurious visual correlations rather than task-relevant causes. We formulate visual-action attribution as an interventional estimation problem. Accordingly, we introduce the Interventional Significance Score (ISS), an interventional masking procedure for estimating the causal influence of visual regions on action predictions, and the Nuisance Mass Ratio (NMR), a scalar measure of attribution to task-irrelevant features. We analyze the statistical properties of ISS and show that it admits unbiased estimation, and we characterize conditions under which action prediction error provides a valid proxy for causal influence. Experiments across diverse manipulation tasks indicate that NMR predicts generalization behavior and that ISS yields more faithful explanations than existing interpretability methods. These results suggest that interventional attribution provides a simple diagnostic approach for identifying causal misalignment in embodied policies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper formulates visual-action attribution in Vision-Language-Action (VLA) policies as an interventional estimation problem. It introduces the Interventional Significance Score (ISS), an interventional masking procedure claimed to estimate causal influence of visual regions on actions with unbiased statistical properties, and the Nuisance Mass Ratio (NMR), a scalar quantifying attribution to task-irrelevant features. The work characterizes conditions under which action prediction error serves as a valid proxy for causal influence, and reports experiments on diverse manipulation tasks showing that NMR predicts generalization behavior while ISS produces more faithful explanations than existing interpretability methods.

Significance. If the unbiased estimation property of ISS and the predictive link from NMR to generalization hold under realistic conditions, the work supplies a practical diagnostic for causal misalignment in embodied policies. This could help identify when VLA models rely on spurious visual correlations, informing more robust training regimes or architectures in robotics. The explicit formulation as interventional estimation and the introduction of NMR as a generalization predictor are concrete contributions that tie interpretability directly to a key failure mode in the field.

major comments (1)
  1. [Abstract / statistical properties of ISS] Abstract and statistical properties section: the claim that ISS 'admits unbiased estimation' and that action prediction error is a valid proxy under 'characterized conditions' does not explicitly bound the degree of feature dependence or account for nonlinear feature extraction in VLA models. In manipulation datasets, visual regions (e.g., gripper pose and object position) are typically correlated; masking one region can implicitly alter encodings of correlated task-relevant features, potentially biasing the interventional estimates even when the stated proxy conditions hold. This directly affects both the NMR-generalization link and the claim of greater faithfulness versus baselines.
minor comments (2)
  1. [Abstract] The abstract and introduction would benefit from a brief explicit statement of the precise proxy conditions (e.g., independence assumptions or error bounds) rather than referring only to 'characterized conditions.'
  2. [Experiments] Experiments section: clarify how the diverse manipulation tasks were selected and whether they include controlled distribution shifts that isolate spurious correlations versus genuine causal features.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thoughtful and constructive feedback on the statistical properties of ISS. The concern regarding unaddressed feature dependence and nonlinear extraction is well-taken and points to a place where the manuscript can be strengthened. We respond to the major comment below and will incorporate revisions to make the assumptions and their limitations more explicit.

read point-by-point responses

  1. Referee: [Abstract / statistical properties of ISS] Abstract and statistical properties section: the claim that ISS 'admits unbiased estimation' and that action prediction error is a valid proxy under 'characterized conditions' does not explicitly bound the degree of feature dependence or account for nonlinear feature extraction in VLA models. In manipulation datasets, visual regions (e.g., gripper pose and object position) are typically correlated; masking one region can implicitly alter encodings of correlated task-relevant features, potentially biasing the interventional estimates even when the stated proxy conditions hold. This directly affects both the NMR-generalization link and the claim of greater faithfulness versus baselines.

    Authors: We appreciate the referee highlighting this subtlety. In the statistical properties section, unbiasedness of ISS is derived by treating the VLA model as a black-box mapping from masked inputs to actions; the intervention is realized by zeroing out selected visual regions and measuring the resulting change in prediction error. Because the error is evaluated after the model's full (nonlinear) feature extraction pipeline, the derivation already incorporates arbitrary nonlinearities in the encoder. However, the stated conditions for the proxy do assume that the masked region is the dominant causal variable and do not explicitly quantify allowable correlations between regions. The referee is correct that, in typical manipulation data, gripper and object features are correlated, so masking one region can indirectly affect the encoding of the other and thereby introduce bias. We will revise the manuscript to (i) add an explicit bound on feature dependence (e.g., a correlation-coefficient threshold under which the bias remains below a stated tolerance), (ii) include a short limitations paragraph discussing the interaction of nonlinear extraction with correlated inputs, and (iii) note the consequent implications for the NMR-generalization correlation and for comparisons against baseline attribution methods. These additions will be placed in the statistical properties section and the experimental discussion without changing the core claims or experimental results. revision: yes

Circularity Check

0 steps flagged

No circularity: definitions and claims remain independent of fitted outputs

full rationale

The paper formulates visual-action attribution as an interventional estimation problem, defines ISS via masking, proves unbiasedness from statistical properties of that procedure, and separately characterizes conditions for the prediction-error proxy. NMR is then defined as a derived scalar from those attributions. None of these steps reduce by construction to the experimental outcomes or to self-referential fitting; the generalization link is presented as an empirical observation rather than a definitional identity. No self-citation chains or ansatz smuggling appear in the provided derivation outline.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The central claim rests on the validity of treating visual-action attribution as an interventional estimation problem and on the statistical properties of the masking procedure; no free parameters are mentioned, but the proxy condition for prediction error is a key domain assumption.

axioms (2)
  • domain assumption Action prediction error provides a valid proxy for causal influence under characterized conditions
    Explicitly stated in the abstract as a condition the authors characterize to support the use of prediction error in attribution.
  • domain assumption Interventional masking yields unbiased estimation of causal influence
    Claimed statistical property of ISS that underpins the method's soundness.
invented entities (2)
  • Interventional Significance Score (ISS) no independent evidence
    purpose: Estimates causal influence of visual regions on action predictions via masking
    Newly introduced measure; no independent evidence provided beyond the abstract's description.
  • Nuisance Mass Ratio (NMR) no independent evidence
    purpose: Scalar measure of attribution to task-irrelevant features
    Newly introduced measure; no independent evidence provided beyond the abstract's description.

pith-pipeline@v0.9.0 · 5460 in / 1437 out tokens · 36033 ms · 2026-05-09T19:41:54.459499+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 12 canonical work pages · 4 internal anchors

  1. [1]

    Forty-Second

    Chen, Letian and Moorman, Nina Marie and Gombolay, Matthew Craig , year =. Forty-Second

  2. [2]

    Forty-Second

    Huang, Huang and Liu, Fangchen and Fu, Letian and Wu, Tingfan and Mukadam, Mustafa and Malik, Jitendra and Goldberg, Ken and Abbeel, Pieter , year =. Forty-Second

  3. [3]

    Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models , booktitle =

    Shi, Lucy Xiaoyang and Equi, Michael Robert and Ke, Liyiming and Pertsch, Karl and Vuong, Quan and Tanner, James and Walling, Anna and Wang, Haohuan and Fusai, Niccolo and. Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models , booktitle =

  4. [4]

    Forty-Second

    Zhang, Hongyin and Zhuang, Zifeng and Zhao, Han and Ding, Pengxiang and Lu, Hongchao and Wang, Donglin , year =. Forty-Second

  5. [5]

    Forty-Second

    Zhang, Jianke and Guo, Yanjiang and Hu, Yucheng and Chen, Xiaoyu and Zhu, Xiang and Chen, Jianyu , year =. Forty-Second

  6. [6]

    Proceedings of the 41st

    Zhen, Haoyu and Qiu, Xiaowen and Chen, Peihao and Yang, Jincheng and Yan, Xin and Du, Yilun and Hong, Yining and Gan, Chuang , year =. Proceedings of the 41st

  7. [7]

    Counterfactual

    Peng, Zhenghao and Ding, Wenhao and You, Yurong and Chen, Yuxiao and Luo, Wenjie and Tian, Thomas and Cao, Yulong and Sharma, Apoorva and Xu, Danfei and Ivanovic, Boris , year =. Counterfactual

  8. [8]

    The 25th International Conference on Autonomous Agents and Multi-Agent Systems , year=

    Don’t Blind Your VLA: Aligning Visual Representations for OOD Generalization , author=. The 25th International Conference on Autonomous Agents and Multi-Agent Systems , year=

  9. [9]

    A Survey on Vision-Language-Action Models: An Action Tokenization Perspective,

    A Survey on Vision-Language-Action Models: An Action Tokenization Perspective , author=. Arxiv Preprint Arxiv:2507.01925 , year=

    work page arXiv

  10. [10]

    Arxiv Preprint Arxiv:2511.18617 , year=

    AutoFocus-IL: VLM-based Saliency Maps for Data-Efficient Visual Imitation Learning without Extra Human Annotations , author=. Arxiv Preprint Arxiv:2511.18617 , year=

    work page arXiv

  11. [11]

    arXiv preprint arXiv:2512.11921 , year =

    Towards Accessible Physical AI: LoRA-Based Fine-Tuning of VLA Models for Real-World Robot Control , author=. Arxiv Preprint Arxiv:2512.11921 , year=

    work page arXiv

  12. [12]

    LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models

    LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models , author=. Arxiv Preprint Arxiv:2510.13626 , year=

    work page internal anchor Pith review arXiv

  13. [13]

    Arxiv Preprint Arxiv:2601.16065 , year=

    DTP: A Simple yet Effective Distracting Token Pruning Framework for Vision-Language Action Models , author=. Arxiv Preprint Arxiv:2601.16065 , year=

    work page arXiv

  14. [14]

    2025 , journal =

    Mechanistic Finetuning of Vision-Language-Action Models via Few-Shot Demonstrations , author =. 2025 , journal =

    2025

  15. [15]

    2022 , journal =

    Wayformer: Motion Forecasting via Simple & Efficient Attention Networks , author =. 2022 , journal =

    2022

  16. [16]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success , author=. Arxiv Preprint Arxiv:2502.19645 , year=

    work page internal anchor Pith review arXiv

  17. [17]

    Arxiv Preprint Arxiv:2505.13888 , year=

    InSpire: Vision-Language-Action Models with Intrinsic Spatial Reasoning , author=. Arxiv Preprint Arxiv:2505.13888 , year=

    work page arXiv

  18. [18]

    Ruijie Zheng and Yongyuan Liang and Shuaiyi Huang and Jianfeng Gao and Hal Daum. Trace. The Thirteenth International Conference on Learning Representations , year=

  19. [19]

    Arxiv Preprint Arxiv:2510.16281 , year=

    Do What You Say: Steering Vision-Language-Action Models via Runtime Reasoning-Action Alignment Verification , author=. Arxiv Preprint Arxiv:2510.16281 , year=

    work page arXiv

  20. [20]

    2022 , journal =

    Do As I Can, Not As I Say: Grounding Language in Robotic Affordances , author =. 2022 , journal =

    2022

  21. [21]

    2025 , journal =

    Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning , author =. 2025 , journal =

    2025

  22. [22]

    2025 , journal =

    Mechanistic Interpretability for Steering Vision-Language-Action Models , author =. 2025 , journal =

    2025

  23. [23]

    2023 , journal =

    VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models , author =. 2023 , journal =

    2023

  24. [24]

    2025 , journal =

    Probing a Vision-Language-Action Model for Symbolic States and Integration into a Cognitive Architecture , author =. 2025 , journal =

    2025

  25. [25]

    Emergent World Representations in

    Molinari, Marco and Nevali, Leonardo and Navani, Saharsha and Younis, Omar G , year =. Emergent World Representations in

  26. [26]

    2024 , journal =

    Words in Motion: Extracting Interpretable Control Vectors for Motion Transformers , author =. 2024 , journal =

    2024

  27. [27]

    Wang, Zhihao and Li, Jianxiong and Zheng, Jinliang and Zhang, Wencong and Liu, Dongxiu and Zheng, Yinan and Niu, Haoyi and Yu, Junzhi and Zhan, Xianyuan , year =

  28. [28]

    Concept-Based Dictionary Learning for Inference-Time Safety in Vision

    Siqi Wen and Shu Yang and Shaopeng Fu and Jingfeng Zhang and Lijie Hu and Di Wang , year=. Concept-Based Dictionary Learning for Inference-Time Safety in Vision

  29. [29]

    Towards Reliable

    Yin, Shiyuan and Bai, Chenjia and Zhang, Zihao and Jin, Junwei and Zhang, Xinxin and Zhang, Chi and Li, Xuelong , year =. Towards Reliable

  30. [30]

    Zhang, Dapeng and Shen, Fei and Zhao, Rui and Chen, Yinda and Zhi, Peng and Li, Chenyang and Zhou, Rui and Zhou, Qingguo , year =

  31. [31]

    Pure Vision Language Action (

    Zhang, Dapeng and Sun, Jin and Hu, Chenghui and Wu, Xiaoyan and Yuan, Zhenlong and Zhou, Rui and Shen, Fei and Zhou, Qingguo , year =. Pure Vision Language Action (

  32. [32]

    , year =

    Zhao, Et Al. , year =. Proceedings of the

  33. [33]

    Robotics: Science and Systems XIX , year=

    RT-1: Robotics Transformer for Real-World Control at Scale , author=. Robotics: Science and Systems XIX , year=

  34. [34]

    Conference on Robot Learning , pages=

    Rt-2: Vision-language-action models transfer web knowledge to robotic control , author=. Conference on Robot Learning , pages=. 2023 , organization=

    2023

  35. [35]

    Moo Jin Kim and Karl Pertsch and Siddharth Karamcheti and Ted Xiao and Ashwin Balakrishna and Suraj Nair and Rafael Rafailov and Ethan P Foster and Pannag R Sanketi and Quan Vuong and Thomas Kollar and Benjamin Burchfiel and Russ Tedrake and Dorsa Sadigh and Sergey Levine and Percy Liang and Chelsea Finn , booktitle=. Open. 2024 , url=

    2024

  36. [36]

    First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024 , year=

    Octo: An Open-Source Generalist Robot Policy , author=. First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024 , year=

    2024

  37. [37]

    $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control

    _0 : A Vision-Language-Action Flow Model for General Robot Control , author=. Arxiv Preprint Arxiv:2410.24164 , year=

    work page internal anchor Pith review arXiv

  38. [38]

    Vision-language-action model with open- world embodied reasoning from pretrained knowledge

    Vision-Language-Action Model with Open-World Embodied Reasoning from Pretrained Knowledge , author=. Arxiv Preprint Arxiv:2505.21906 , year=

    work page arXiv

  39. [39]

    Advances in neural information processing systems , volume=

    Decision Transformer: Reinforcement Learning via Sequence Modeling , author=. Advances in neural information processing systems , volume=

  40. [40]

    2014 , publisher=

    Probabilistic reasoning in intelligent systems: networks of plausible inference , author=. 2014 , publisher=

    2014

  41. [41]

    1988 , publisher=

    Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference , author=. 1988 , publisher=

    1988

  42. [42]

    Probabilistic and causal inference: The works of Judea Pearl , pages=

    Equivalence and Synthesis of Causal Models , author=. Probabilistic and causal inference: The works of Judea Pearl , pages=. 2022 , publisher =

    2022

  43. [43]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Exploring the Limits of Vision-Language-Action Manipulation in Cross-task Generalization , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  44. [44]

    International Conference on Machine Learning , pages=

    Certified Adversarial Robustness via Randomized Smoothing , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  45. [45]

    Advances in neural information processing systems , volume=

    Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers , author=. Advances in neural information processing systems , volume=

  46. [46]

    IEEE Robotics and Automation Letters , volume=

    Rlbench: The robot learning benchmark & learning environment , author=. IEEE Robotics and Automation Letters , volume=. 2020 , publisher=

    2020

  47. [47]

    Causality , publisher=

    Pearl, Judea , year=. Causality , publisher=

  48. [48]

    The Do-Calculus Revisited , booktitle =

    Judea Pearl , editor =. The Do-Calculus Revisited , booktitle =

  49. [49]

    Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascal, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas , title =

  50. [50]

    RISE: Randomized Input Sampling for Explanation of Black-box Models

    Rise: Randomized input sampling for explanation of black-box models , author=. Arxiv Preprint Arxiv:1806.07421 , year=

    work page Pith review arXiv

  51. [51]

    2016 , publisher=

    Information geometry and its applications , author=. 2016 , publisher=

    2016

  52. [52]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Sigmoid loss for language image pre-training , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  53. [53]

    PaliGemma: A versatile 3B VLM for transfer

    Paligemma: A versatile 3b vlm for transfer , author=. arXiv preprint arXiv:2407.07726 , year=

    work page internal anchor Pith review arXiv

  54. [54]

    Neural computation , volume=

    A learning algorithm for continually running fully recurrent neural networks , author=. Neural computation , volume=. 1989 , publisher=

    1989