Pith. sign in

REVIEW 4 major objections 7 minor 34 references

ConditionNET: Learning Preconditions and Effects for Execution Monitoring

T0 review · 4 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ConditionNET claims that a 30M-parameter vision-language model can learn robot action preconditions and effects from demonstrations and use them to detect execution anomalies in real time.

desk verdict ConditionNET's three-way precondition/effect/unsatisfied formulation and consistency loss are a genuine step beyond success-detection baselines, but the anomaly-detection evaluation needs a clearer protocol before the headline numbers can be fully trusted. read the letter →

arxiv 2502.01167 v1 pith:47KGLHZN submitted 2025-02-03 cs.RO cs.LG

Daniel Sliwowski , Dongheui Lee This is my paper · ORCID
classification cs.ROcs.LG
keywords ConditionNETpreconditionlearningeffectexecutionmonitoringanomalydetectionvision-languagemodelphasepredictionrobotmanipulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a small vision-language model can learn, from teleoperated demonstrations alone, what must be true before a robot action starts (precondition) and what should be true after it ends (effect), without hand-written planners. It frames each camera frame and action description as a three-way classification: precondition, effect, or neither. On two manipulation datasets the model reports higher accuracy than three baselines, and a real-robot test shows the learned conditions can drive a behavior tree that halts or retries actions when the scene deviates from expectations. The reason to care is that execution monitoring currently relies on hand-coded predicates or heavyweight vision-language models, and ConditionNET is small enough and fast enough to run per frame.

What carries the argument

The load-bearing mechanism is a two-stage transformer built on frozen image features and a frozen text encoder. The first stage (State Transformer) produces a global scene representation; the second stage (Condition Transformer) replaces the class token with the action's language embedding so the model attends to action-relevant objects. The training also imposes a consistency objective: for successful demonstrations, the difference between the effect-state and pre-state feature vectors is pulled, via an InfoNCE loss, toward the semantic embedding of a paraphrased action description. This 'action as state difference' identity is what makes the model learn conditions rather than just memorizing appearance.

What would settle it

Take a set of teleoperated demonstrations, have several people independently annotate the pre/core/post segment boundaries, and measure inter-annotator agreement; if the boundaries vary widely, then the reported phase-prediction accuracy is partly measuring annotation consistency rather than a learned property of actions.

Watch

Extended reading notes

Core claim

ConditionNET's central claim is that the preconditions and effects of manipulation actions can be learned directly from image-action pairs, and that anomalies can be detected by comparing the phase a behavior tree expects with the phase the model predicts. The model classifies each observation as precondition, effect, or unsatisfied; during execution, if the expected phase is precondition but the model says otherwise, the robot halts, and if the expected phase is effect but the model disagrees, the action is retried. The paper reports that ConditionNET outperforms all tested baselines on both anomaly detection and phase prediction, with accuracy 0.97 on its own (Im)PerfectPour dataset versus 0.86 for the best baseline, and that it runs in about 21 ms per batch of six actions with roughly 30 million parameters.

Load-bearing premise

The method's training labels come from human annotations that split each demonstration into preparation, core, and post phases and assume one action's effect segment overlaps the next action's pre-segment; if those boundaries are inconsistent or that overlap fails, the learned preconditions and effects, and the anomaly detector built on them, inherit the error.

Editorial extensions

If this is right

  • Anomaly detection no longer needs hand-written planning predicates; preconditions and effects can be learned from demonstrations with natural language action labels.
  • Because the model is small and fast, it can monitor every frame during execution, so failures such as a bottle being removed before pickup can be caught during the pre-phase rather than after the action completes.
  • The consistency loss improves both phase prediction and anomaly detection, and paraphrased action text during training makes the model robust to varied ways of describing the same action.
  • With fewer than 30 million parameters and about 21 ms inference, the approach is deployable on real robots in a way that billion-parameter vision-language models are not.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the same 'state difference equals action semantics' identity could be used in reverse: given two observations, predict which action was performed, turning the model into a self-supervised action classifier without paired action labels.
  • The overlap assumption between one action's effect segment and the next action's pre-segment suggests the learned conditions are relational, so a natural extension is to chain conditions across multi-step tasks and detect task-level anomalies, not just action-level ones.
  • The reported margin over baselines may shrink if the baselines are given the same two-stage conditioning, so a direct test is to replace the simple MLP head with a transformer conditioned on action text.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes ConditionNET, a compact vision-language model (~30M parameters) that classifies an image-action pair into one of three states: precondition, effect, or unsatisfied. The model is trained on two manipulation datasets—the existing FAILURE dataset and a newly collected (Im)PerfectPour dataset of teleoperated Franka Panda demonstrations—using a cross-entropy phase classification loss and a consistency loss that aligns the difference between effect and precondition features with paraphrased action text embeddings. Anomaly detection is performed by comparing the model's predicted phase with the expected phase supplied by a behavior tree. The paper reports that ConditionNET outperforms all baselines on both anomaly detection and phase prediction, and demonstrates a real-robot execution monitoring system with qualitative results.

Significance. If the evaluation is confirmed, this is a worthwhile contribution: it introduces a data-driven way to learn action preconditions and effects from vision and language, with a model small enough for real-time use, and it provides a new failure-inclusive dataset and a real-robot demonstration. The two-stage transformer architecture and the consistency loss are reasonable design choices. The main weakness is that the quantitative evaluation, especially for anomaly detection, is under-specified; the central claim of outperforming baselines by large margins is not yet verifiable as written.

major comments (4)
  1. [IV-C, Table I] The anomaly detection metrics in Table I are reported without defining the evaluation unit or the ground-truth mapping. It is unclear whether accuracy, precision, recall, and F1 are computed per frame, per action, or per demonstration, and how the human-annotated anomalies described in Section IV-A are converted into binary ground-truth labels at that unit. The 8-frame filtering step in Section III-F must also be applied consistently to all models; FinoNET outputs one label per 8-frame window, so comparing it with ConditionNET's per-frame phase-mismatch predictions is not apples-to-apples as written. Please specify the complete evaluation protocol so the reported margins can be reproduced.
  2. [III-A and IV-A] The construction of the "unsatisfied" class is not described. The model is trained to classify frames as precondition, effect, or unsatisfied, but Section IV-A only mentions annotating pre-, core-, and post-segments. If core-phase frames are assigned the "unsatisfied" label, state this explicitly; otherwise provide the labeling rule for the third class. Without this, the phase prediction task is not fully specified, and the anomaly detection logic that depends on the "unsatisfied" prediction cannot be interpreted.
  3. [IV-C] The quantitative results are reported for a single 70/30 split, without variance over random seeds or cross-validation folds. Given that (Im)PerfectPour has 544 demonstrations and FAILURE has 235, the reported margins (e.g., 7 points in F1 on FAILURE) could be within expected noise. Please report means and standard deviations over at least three random splits or seeds, or use cross-validation, to support the claim of consistent improvement.
  4. [III-F and IV-C] The relation between the anomaly detection metric and the phase prediction metric needs clarification. The anomaly detection procedure compares the predicted phase with an expected phase supplied by the behavior tree; in the offline evaluation, the expected phase presumably comes from the same temporal annotations used as training labels. If so, a frame is flagged as anomalous exactly when the predicted phase disagrees with the expected phase on pre/post segments, making the anomaly detection metric a re-encoding of phase prediction accuracy. Please describe the exact offline computation of Table I's anomaly detection scores and discuss how this differs from the success/failure task solved by FinoNET. This is necessary to establish that the comparison in Table I is meaningful and that the model detects semantically meaningful anomalies beyond phase misclassification.
minor comments (7)
  1. [IV-C] The claim of "a margin of at least 10% for the anomaly detection task" is not consistently supported if F1 is considered: on FAILURE the F1 margin over the best baseline (CLIP+MLP) is 7 points. Clarify which metric and which baseline are used for this claim, and reconcile it with "at least 6%" in the conclusion.
  2. [III-B] The data augmentation mixing ratio of 0.5 is described as "experimentally determined"; please report whether this was tuned on the validation set and include a brief sensitivity analysis or at least state the range tested.
  3. [IV-A] Please clarify whether demonstrations recorded with two cameras are counted as separate demonstrations in the dataset statistics and in the evaluation split.
  4. [IV-C] The inference time is reported as 21 ± 18 ms; with such a large standard deviation, the mean is not very informative. Report the median or percentile values as well.
  5. [III-D] The indicator function in Eq. (2) is explained only as filtering out unsuccessful demonstrations; specify how it is computed for each triplet, since a demonstration-level failure label may not map cleanly to individual image pairs.
  6. [Table I] Define the abbreviations "Pre" and "Rec" in the table caption (presumably precision and recall).
  7. [II-B and III-B] Use consistent naming for the models: "PALM-E" is spelled "PALM" in one place, and "ChatGPT-4" in Section IV-C differs from "ChatGPT-4o" in Section III-B.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ConditionNET's anomaly detection is a defined post-processing of phase predictions, evaluated on a held-out split.

full rationale

ConditionNET's central derivation is self-contained. The model is trained with a cross-entropy phase-classification loss (Eq. 1) and a consistency regularizer (Eq. 2); neither objective fits the anomaly-detection metric. The claimed anomaly-detection performance is obtained by comparing the model's predicted phase to the expected phase from a behavior tree (Sec. III-F), which is an external specification, not a model output. Evaluation uses a 70/30 held-out split (Sec. IV-C), and hyperparameters such as the 0.5 augmentation ratio and the 8-frame smoothing filter are standard tuning choices, not fitted predictions. The paper includes self-citations (e.g., [3]–[5]) but only for contextual discussion of planning and recovery; none is load-bearing for the central learning claim. The only notable weakness is that the anomaly-detection evaluation unit and ground-truth mapping are not fully specified in Sec. IV-C, which is a reporting gap rather than circularity. No equation reduces to a fitted value, and no claimed result is forced by a self-citation chain.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on human-provided phase segmentation, an overlap assumption between effect and pre segments, and the visual distinguishability of anomalies. The method introduces no new physical entities; the learned condition feature is a model component. Free parameters are limited to validation-tuned data augmentation and filtering choices.

free parameters (3)
  • data augmentation mixing ratio = 0.5
    Selected by experiment as yielding best performance (Section III-B).
  • anomaly filtering window = 8 consecutive frames
    Introduced to suppress false positive anomaly detections; threshold chosen by authors (Section III-F).
  • paraphrase count = 20 variants per action/object
    Number of LLM-generated action descriptions used for augmentation (Section III-B).
assumptions (6)
  • domain assumption Human annotations of preparation/core/post segments and success labels are accurate enough to supervise condition learning.
    The temporal segmentation from Section IV-A provides all ground-truth phase labels for the three-way classification; errors propagate into learned conditions.
  • domain assumption The effect segment of one action overlaps the pre-segment of the next action.
    Explicit labeling assumption in Section IV-A; used to define what counts as an effect and to relate actions.
  • domain assumption The pre-state of one action can be used as the post-state of a related action for data augmentation.
    Data augmentation in Section III-B relies on known relationships between actions in the dataset, e.g., pre of pick equals post of place.
  • domain assumption The robot is the only agent affecting the scene and anomalies are visually distinguishable.
    Dataset construction assumption in Section IV-A; excludes non-visual failures and concurrent human interventions.
  • domain assumption Core motion phase states are ambiguous, so anomaly detection is suspended during that phase.
    Design choice in Section III-F; means anomalies occurring during the core phase are not detected until the expected phase changes.
  • domain assumption The vector difference between effect-state and pre-state features should be alignable with the action description.
    Central to the consistency loss in Section III-D; if the geometry does not hold, the regularizer may not improve representation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ConditionNET: Learning Preconditions and Effects for Execution Monitoring." pith.science (2026). https://pith.science/paper/47KGLHZN

@misc{pith2026250201167,
  author       = {Pith},
  title        = {Pith review of: ConditionNET: Learning Preconditions and Effects for Execution Monitoring},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/47KGLHZN}},
  note         = {Machine review of arXiv:2502.01167}
}
read the original abstract

The introduction of robots into everyday scenarios necessitates algorithms capable of monitoring the execution of tasks. In this paper, we propose ConditionNET, an approach for learning the preconditions and effects of actions in a fully data-driven manner. We develop an efficient vision-language model and introduce additional optimization objectives during training to optimize for consistent feature representations. ConditionNET explicitly models the dependencies between actions, preconditions, and effects, leading to improved performance. We evaluate our model on two robotic datasets, one of which we collected for this paper, containing 406 successful and 138 failed teleoperated demonstrations of a Franka Emika Panda robot performing tasks like pouring and cleaning the counter. We show in our experiments that ConditionNET outperforms all baselines on both anomaly detection and phase prediction tasks. Furthermore, we implement an action monitoring system on a real robot to demonstrate the practical applicability of the learned preconditions and effects. Our results highlight the potential of ConditionNET for enhancing the reliability and adaptability of robots in real-world environments. The data is available on the project website: https://dsliwowski1.github.io/ConditionNET_page.

Figures

Figures reproduced from arXiv: 2502.01167 by the authors.

Figure 1
Figure 1. An overview of the proposed anomaly detection and recovery algorithm. ConditionNET detects anomalies by comparing the expected and current motion phases, the latter is predicted by a vision-language model. A behavior tree governs task execution, providing the action and expected phases, while motions are generated by a skill library and executed on the robot with an impedance controller. Many studies focus on execut… view at source ↗
Figure 2
Figure 2. ConditionNET Architecture. For an image-action pair, we compute the condition feature E and classify the current observation as precondition, effect, or unsatisfied. We extract image and semantic features using DINOv2 [27] and CLIP [8]. The State Transformer then extracts the general state feature cls c, and the Condition Transformer extracts the condition feature E. For consistency loss, we use features from both t… view at source ↗
Figure 3
Figure 3. Simplified depiction of the Behavior tree. Octagons – Selector nodes, rectangles – sequence nodes, rounded rect￾angles – behaviors. Lconsistency = − 1 B X B i=1 1(i) log exp(Sii/τ ) PB k=1 1(k) exp(Sik/τ ) − 1 B X B j=1 1(j) log exp(Sjj/τ ) PB k=1 1(k) exp(Skj/τ ) , (2) Sij = e i a ◦ s j p ∥e i a∥∥s j p∥ , (3) where Sij is the cosine similarity between feature i and feature j, τ is a learned temperature parameter, a… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative Results show the model performance in the continuous action monitoring experiment. For clarity of presentation, only the results for single actions have been shown, but in reality, predictions for all actions are made in parallel. We highlight less visible …
Figure 5
Figure 5. Figure 5: Phase prediction confidences and anomaly prediction results over time. Each hue marks a different expected action, and [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages

  1. [1]

    A review of robot learning for manipulation: Challenges, representations, and algorithms,

    O. Kroemer, S. Niekum, and G. Konidaris, “A review of robot learning for manipulation: Challenges, representations, and algorithms,” J. Mach. Learn. Res. , vol. 22, no. 1, jan 2021

    work page 2021

  2. [2]

    Grounding classical task planners via vision-language models,

    X. Zhang, Y . Ding, S. Amiri, H. Yang, A. Kaminski, C. Esselink, and S. Zhang, “Grounding classical task planners via vision-language models,” in arXiv preprint arXiv:2304.08587 , 04 2023

    arXiv 2023

  3. [3]

    Long-Horizon Planning and Execution with Functional Object-Oriented Networks,

    D. Paulius, A. Agostini, and D. Lee, “Long-Horizon Planning and Execution with Functional Object-Oriented Networks,” IEEE Robotics and Automation Letters , vol. 8, no. 8, pp. 4513–4520, 2023

    work page 2023

  4. [4]

    Multi-level task learning based on intention and constraint inference for autonomous robotic manipulation,

    C. Willibald and D. Lee, “Multi-level task learning based on intention and constraint inference for autonomous robotic manipulation,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 7688–7695

    work page 2022

  5. [5]

    Intuitive programming of conditional tasks by demonstration of multiple solutions,

    T. Eiband, M. Saveriano, and D. Lee, “Intuitive programming of conditional tasks by demonstration of multiple solutions,” IEEE Robotics and Automation Letters , vol. 4, no. 4, pp. 4483–4490, 2019

    work page 2019

  6. [6]

    Representing robot task plans as robust logical-dynamical systems,

    C. Paxton, N. D. Ratliff, C. Eppner, and D. Fox, “Representing robot task plans as robust logical-dynamical systems,” 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pp. 5588–5595, 2019. [Online]. Available: https://api.semanticscholar.org/ CorpusID:199453145

    work page 2019

  7. [7]

    Pddl - the planning domain definition language,

    G. M. et al., “Pddl - the planning domain definition language,” Techincal Report, 08 1998

    work page 1998

  8. [8]

    Learning transferable visual models from natural language supervision,

    A. Radford et al., “Learning transferable visual models from natural language supervision,” in Proceedings of the 38th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, M. Meila and T. Zhang, Eds., vol. 139. PMLR, 18–24 Jul 2021, pp. 8748–8763. [Online]. Available: https://proceedings.mlr.press/v139/radford21a.html

    work page 2021

Show all 34 references
  1. [9]

    Instructblip: Towards general-purpose vision-language models with instruction tuning,

    W. D. et al., “Instructblip: Towards general-purpose vision-language models with instruction tuning,” 2023

    2023

  2. [10]

    The safety filter: A unified view of safety-critical control in autonomous systems,

    K.-C. Hsu, H. Hu, and J. Fisac, “The safety filter: A unified view of safety-critical control in autonomous systems,” Annual Review of Control, Robotics, and Autonomous Systems , vol. 7, 02 2024

    2024

  3. [11]

    Collaborative programming of conditional robot tasks,

    C. Willibald, T. Eiband, and D. Lee, “Collaborative programming of conditional robot tasks,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2020, pp. 5402–5409

    2020

  4. [12]

    Anomaly detection for insertion tasks in robotic assembly using gaus- sian process models,

    D. Romeres, D. K. Jha, W. Yerazunis, D. Nikovski, and H. A. Dau, “Anomaly detection for insertion tasks in robotic assembly using gaus- sian process models,” in 2019 18th European Control Conference (ECC). IEEE, 2019, pp. 1017–1022

    2019

  5. [13]

    Hmms for anomaly detection in autonomous robots,

    D. Azzalini, A. Castellini, M. Luperto, A. Farinelli, and F. Amigoni, “Hmms for anomaly detection in autonomous robots,” in Int. Conf. on Autonomous Agents and MultiAgent Systems , 2020, pp. 105–113

    2020

  6. [14]

    Multimodal anomaly detection for assistive robots,

    D. Park, H. Kim, and C. C. Kemp, “Multimodal anomaly detection for assistive robots,” Autonomous Robots , vol. 43(3), pp. 611–629, 2019

    2019

  7. [15]

    Regularized parameter estimation in high-dimensional gaussian mixture models,

    L. Ruan, M. Yuan, and H. Zou, “Regularized parameter estimation in high-dimensional gaussian mixture models,” Neural computation , vol. 23, no. 6, pp. 1605–1622, 2011

    2011

  8. [16]

    Multimodal anomaly detection based on deep auto-encoder for object slip perception of mobile ma- nipulation robots,

    Y . Yoo, C.-Y . Lee, and B.-T. Zhang, “Multimodal anomaly detection based on deep auto-encoder for object slip perception of mobile ma- nipulation robots,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021, pp. 11 443–11 449

    2021

  9. [17]

    Multimodal detection and clas- sification of robot manipulation failures,

    A. Inceoglu, E. E. Aksoy, and S. Sariel, “Multimodal detection and clas- sification of robot manipulation failures,” IEEE Robotics and Automation Letters, vol. 9, no. 2, pp. 1396–1403, 2024

    2024

  10. [18]

    Clue-ai: A convolutional three-stream anomaly identification framework for robot manipulation,

    D. Altan and S. Sariel, “Clue-ai: A convolutional three-stream anomaly identification framework for robot manipulation,” IEEE Access, vol. 11, pp. 48 347–48 357, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:247476170

    2022

  11. [19]

    Fino-net: A deep multimodal sensor fusion framework for manipulation failure detection,

    A. Inceoglu, E. E. Aksoy, A. Cihan Ak, and S. Sariel, “Fino-net: A deep multimodal sensor fusion framework for manipulation failure detection,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 6841–6847

    2021

  12. [20]

    Inner monologue: Embodied reasoning through planning with language models,

    W. Huang et al., “Inner monologue: Embodied reasoning through planning with language models,” in Proceedings of The 6th Conference on Robot Learning , ser. Proceedings of Machine Learning Research, K. Liu, D. Kulic, and J. Ichnowski, Eds., vol. 205. PMLR, 14–18 Dec 2023, pp. 1...

    2023

  13. [21]

    Vision-language models as success detectors,

    Y . D. et al., “Vision-language models as success detectors,” ArXiv, vol. abs/2303.07280, 2023. [Online]. Available: https://api.semanticscholar. org/CorpusID:257496810

    2023 arXiv

  14. [22]

    REFLECT: Summarizing robot experiences for failure explanation and correction,

    Z. Liu, A. Bahety, and S. Song, “REFLECT: Summarizing robot experiences for failure explanation and correction,” in 7th Annual Conference on Robot Learning , 2023. [Online]. Available: https: //openreview.net/forum?id=8yTS nAILxt

    2023

  15. [23]

    A survey of behavior trees in robotics and ai,

    M. I. et al., “A survey of behavior trees in robotics and ai,” Robotics and Autonomous Systems , vol. 154, p. 104096, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0921889022000513

    2022

  16. [24]

    Learning strips action models with classical planning,

    D. Aineto, S. Jim ´enez, and E. Onaind ´ıa, “Learning strips action models with classical planning,” in International Conference on Automated Planning and Scheduling , 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:49405691

    2018

  17. [25]

    Learning symbolic operators for task and motion planning,

    T. S. et al., “Learning symbolic operators for task and motion planning,” 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pp. 3182–3189, 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:232076417

    2021

  18. [26]

    Palm-e: An embodied multimodal language model,

    D. Driess et al., “Palm-e: An embodied multimodal language model,” in arXiv preprint arXiv:2303.03378 , 2023

    2023 arXiv

  19. [27]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab et al., “Dinov2: Learning robust visual features without supervision,” 2023

    2023

  20. [28]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv preprint arXiv:1807.03748 , 2018

    2018 arXiv

  21. [29]

    ChatGPT (May 16 version ),

    OpenAI, “ChatGPT (May 16 version ),” https://chat.openai.com/chat

  22. [30]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  23. [31]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    Alexey Dosovitskiy et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” in International Conference on Learning Representations , 2021. [Online]. Available: https://openreview.net/forum?id=YicbFdNTTy

    2021

  24. [32]

    Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks,

    J. L. et al., “Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks,” in Neural Information Processing Systems , 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:199453025

    2019

  25. [33]

    AllenNLP: A deep semantic natural language processing platform,

    M. e. a. Gardner, “AllenNLP: A deep semantic natural language processing platform,” in Proceedings of Workshop for NLP Open Source Software (NLP-OSS) , E. L. Park, M. Hagiwara, D. Milajevs, and L. Tan, Eds. Melbourne, Australia: Association for Computational Linguistics, Jul. ...

    2018

  26. [34]

    Openflamingo: An open-source framework for training large autoregressive vision-language models,

    Anas Awadalla et al., “Openflamingo: An open-source framework for training large autoregressive vision-language models,” arXiv preprint arXiv:2308.01390, 2023

    2023 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.