pith. sign in

arxiv: 2606.23420 · v1 · pith:ZEITMKBRnew · submitted 2026-06-22 · 💻 cs.RO

Flowing With Purpose: Latent Action Guided Flow Matching Policies For Robotic Manipulation

Bruno Machado , Alexandre Chapin , Emmanuel Dellandrea , Liming Chen This is my paper

Pith reviewed 2026-06-26 08:00 UTC · model grok-4.3

classification 💻 cs.RO
keywords flow matchingrobotic manipulationlatent action modelbehavior cloningpolicy learninggenerative modelingaction primitives
0
0 comments X

The pith

Replacing fixed Gaussian sources with adaptive priors from latent action models improves robotic flow matching policies.

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

The paper establishes that standard flow matching policies for robotic manipulation are limited by their use of a single fixed isotropic Gaussian source distribution, which leads to entangled vector fields due to the fragmented and heteroscedastic nature of action spaces. LAFM addresses this by using a latent action model to select specialized prior distributions that align with observed motion primitives, making the denoising process more efficient. This results in higher success rates in both simulated benchmarks and real robot deployments, even outperforming larger pre-trained models with smaller networks. A reader would care because it offers a targeted fix to a structural issue in generative policy learning without requiring more data or compute.

Core claim

By grounding an adaptive library of learned prior distributions with a latent action model that maps observations to discrete motion primitives, LAFM provides structurally aligned initializations for the flow matching process, producing shorter and less entangled transport trajectories that accommodate the heteroscedasticity in human demonstrations.

What carries the argument

The latent action model that selects specialized base distributions from an adaptive library to initialize flow matching denoising.

Load-bearing premise

The globally fixed isotropic source distribution is the main cause of entangled vector fields in robotic flow matching, and switching to latent-action-selected adaptive priors will resolve this without introducing training instabilities.

What would settle it

A direct comparison of learned vector field entanglement metrics or transport path lengths between standard flow matching and LAFM on the same robotic tasks, where no improvement would falsify the benefit of the adaptive initialization.

Figures

Figures reproduced from arXiv: 2606.23420 by Alexandre Chapin, Bruno Machado, Emmanuel Dellandrea, Liming Chen.

Figure 1
Figure 1. Figure 1: LAFM overview. Left: We train a LAM to extract latent actions from consecutive frames. Right: We divide our policy into two parts: The encoder predicts the next latent action for a given frame. The predicted latent action is used to condition the noise sampling, guiding the flow matching process in the decoder. This limitation is further amplified by the heteroscedastic nature of manipulation demonstration… view at source ↗
Figure 2
Figure 2. Figure 2: Implemented architecture. The images, task description, and robot’s proprioception are given to the policy encoder. The images and text are first processed by a specialized encoder before being fed to a Transformer. The output of the Transformer encoder is given via cross-attention to the DiT decoder. The latent action predicted by the encoder is used to choose the distribution from which the noise is samp… view at source ↗
Figure 3
Figure 3. Figure 3: Learned prior distributions visualization. We plot the t-SNE of the learned means. The learned variances are represented by the color and size of the circles. We display the optical flow generated by the LAM decoder while reconstructing the same image using different latent actions. PC1 (72.3%) PC2 (12.7%) Flow Matching L1 Error: 0.54 Transport Cost: 139.7 PC1 (72.3%) Latent Action Conditioned Flow Matchin… view at source ↗
Figure 4
Figure 4. Figure 4: Vector fields visualization. We compare the denoising process of the standard and latent action guided flow matching policies. The gray dots are the sampled noise points, the black dots are the final predicted actions, and the blue lines connecting the dots represents the flow matching vector fields. The target actions are represented by the red crosses, being the same on both plots. We present the average… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of different LAM design choices on the final policy performance. We ablate the [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation on the number of denoising steps. We plot the success rate on LIBERO-90 for LAFM and FM policies using different numbers of denoising steps during inference. Flow matching step sampling. During training, we follow Black et al. [2] and sample the flow matching step τ from a scaled Beta distribution that prioritizes noisier samples and adds a cutoff value of 0.999. Specifically: u ∼ Beta(1, 1.5), τ … view at source ↗
Figure 7
Figure 7. Figure 7: illustrates the four real-world tasks proposed for our experiments [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
read the original abstract

Flow matching has recently become a new standard for behavior cloning in robotic manipulation. However, state-of-the-art flow matching policies suffer from a systematic structural mismatch: they rely on a globally fixed isotropic source distribution despite the strongly fragmented and heteroscedastic structure of robotic action spaces. This agnostic initialization forces the model to learn highly entangled vector fields, bottlenecking training efficiency and limiting overall policy performance. To address this limitation, we introduce Latent Action Guided Flow Matching (LAFM), a novel framework that replaces the monolithic Gaussian with an adaptive library of learned prior distributions. By grounding these distributions using a latent action model, LAFM maps current observations to discrete motion primitives, selecting a specialized base distribution that provides an informed, structurally aligned initialization for the denoising process. This dynamic adaptivity naturally accommodates heteroscedasticity in human demonstrations and makes transport trajectories shorter and less entangled. Empirically, LAFM substantially outperforms standard flow matching formulations, increasing task success rates by 23.4% in real-world robotic deployments and by 10.4% on the LIBERO-90 benchmark. Furthermore, we demonstrate that LAFM achieves state-of-the-art results, surpassing massively pre-trained vision-language-action models while utilizing significantly smaller architectures.

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

2 major / 1 minor

Summary. The manuscript proposes Latent Action Guided Flow Matching (LAFM) to address a claimed structural limitation in flow matching policies for robotic manipulation. Standard methods use a fixed isotropic Gaussian source distribution, which the authors argue forces entangled vector fields in heteroscedastic action spaces. LAFM replaces this with an adaptive library of learned priors selected via a latent action model that maps observations to discrete motion primitives. The paper reports empirical gains of 23.4% task success in real-world deployments and 10.4% on LIBERO-90, along with state-of-the-art results using smaller architectures than massively pre-trained vision-language-action models.

Significance. If the performance claims hold under rigorous controls, the work could meaningfully advance flow-matching approaches in robotics by providing a mechanism for structurally aligned initialization that accommodates action-space heterogeneity. The idea of grounding priors in latent motion primitives is a plausible direction for improving sample efficiency and reducing training entanglement, and the reported ability to outperform larger models with compact architectures would be a notable practical contribution if the comparisons are equitable.

major comments (2)
  1. [Abstract] Abstract: The central performance claims (23.4% real-world and 10.4% LIBERO-90 gains, plus SOTA over larger VLA models) are presented without any description of experimental protocol, baselines, number of trials, statistical tests, or ablation studies. This absence makes the empirical contribution impossible to evaluate and is load-bearing for the paper's primary assertion.
  2. [Abstract] Abstract: No technical details are supplied on the latent action model architecture, its training objective, how the library of priors is constructed or selected, or how the selection integrates into the flow-matching ODE. Without these elements it is impossible to assess whether the adaptive priors actually shorten trajectories or avoid new instabilities, undermining evaluation of the core methodological claim.
minor comments (1)
  1. [Abstract] Abstract: The phrase 'massively pre-trained vision-language-action models' and 'significantly smaller architectures' are used without naming the specific models, parameter counts, or training regimes, which weakens the interpretability of the size-performance comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful comments. The two major points both concern the level of detail provided in the abstract. We address them point-by-point below, noting that abstracts are intentionally concise summaries while the full experimental and methodological details appear in the body of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claims (23.4% real-world and 10.4% LIBERO-90 gains, plus SOTA over larger VLA models) are presented without any description of experimental protocol, baselines, number of trials, statistical tests, or ablation studies. This absence makes the empirical contribution impossible to evaluate and is load-bearing for the paper's primary assertion.

    Authors: We agree that the abstract itself contains no experimental protocol, baselines, trial counts, statistical tests, or ablations; this is standard for abstracts due to length limits. The full manuscript supplies these details in Section 4 (Experiments), including the real-world deployment protocol (number of trials per task, success criteria), LIBERO-90 evaluation setup, baseline implementations (standard flow matching and large VLA models), statistical reporting, and ablation studies on the latent action model. The performance numbers are therefore supported by the complete experimental record rather than the abstract alone. We do not believe a change to the abstract is required for the claims to be evaluable. revision: no

  2. Referee: [Abstract] Abstract: No technical details are supplied on the latent action model architecture, its training objective, how the library of priors is constructed or selected, or how the selection integrates into the flow-matching ODE. Without these elements it is impossible to assess whether the adaptive priors actually shorten trajectories or avoid new instabilities, undermining evaluation of the core methodological claim.

    Authors: The abstract likewise omits these architectural and integration specifics for brevity. Section 3 (Method) fully specifies the latent action model (architecture, training objective for learning discrete motion primitives), the construction and selection mechanism for the library of priors, and its integration into the flow-matching ODE (observation-conditioned selection of the base distribution). The text explains how this choice produces shorter, less entangled transport trajectories. The core methodological claim can therefore be assessed from the full manuscript; we see no need to expand the abstract. revision: no

Circularity Check

0 steps flagged

No circularity detected; empirical method with independent performance claims

full rationale

The provided abstract and context describe an empirical framework (LAFM) that replaces a fixed isotropic Gaussian source in flow matching with adaptive priors selected via a latent action model. No equations, derivation steps, or self-citations appear in the text. The central claims are performance improvements measured on benchmarks and real-world tasks, which do not reduce to fitted inputs by construction or rely on self-referential definitions. The motivation (heteroscedasticity mismatch) is stated as an assumption but is not used to derive results tautologically. This is a standard empirical contribution without load-bearing circular steps.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Only the abstract is available, so the ledger reflects assumptions and entities explicitly named or implied there; full details on parameters and training are absent.

free parameters (1)
  • parameters of the latent action model
    The model that maps observations to discrete motion primitives must be trained on demonstration data, introducing fitted parameters whose values are not reported.
axioms (1)
  • domain assumption Robotic action spaces exhibit strongly fragmented and heteroscedastic structure that a single isotropic Gaussian cannot capture efficiently.
    Directly stated in the abstract as the motivation for replacing the fixed source distribution.
invented entities (1)
  • adaptive library of learned prior distributions no independent evidence
    purpose: To supply observation-conditioned, structurally aligned initializations for the flow matching denoising process.
    Introduced in the abstract as the core replacement for the monolithic Gaussian; no independent evidence outside the paper is provided.

pith-pipeline@v0.9.1-grok · 5753 in / 1370 out tokens · 37027 ms · 2026-06-26T08:00:50.895913+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

37 extracted references · 9 canonical work pages

  1. [1]

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

    Cheng Chi et al. “Diffusion Policy: Visuomotor Policy Learning via Action Diffusion”. In: Proceedings of Robotics: Science and Systems. Daegu, Republic of Korea, July 2023.DOI: 10.15607/RSS.2023.XIX.026(cit. on pp. 1, 6–8)

    work page doi:10.15607/rss.2023.xix.026(cit 2023

  2. [2]

    π0: A Vision-Language-Action Flow Model for General Robot Control

    Kevin Black et al. π0: A Vision-Language-Action Flow Model for General Robot Control

  3. [3]

    arXiv: 2410.24164 [cs.LG].URL: https://arxiv.org/abs/2410.24164 (cit. on pp. 1, 3, 6, 7, 9, 16, 17)

    Pith/arXiv arXiv

  4. [4]

    NVIDIA et al.GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. 2025. arXiv: 2503.14734 [cs.RO].URL: https://arxiv.org/abs/2503.14734 (cit. on pp. 1, 9)

    Pith/arXiv arXiv 2025

  5. [5]

    Denoising Diffusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising Diffusion Probabilistic Models”. In: Advances in Neural Information Processing Systems. Ed. by H. Larochelle et al. V ol. 33. Curran Associates, Inc., 2020, pp. 6840–6851.URL: https://proceedings.neurips.cc/ paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf (cit. on p. 1)

    2020

  6. [6]

    Deep Unsupervised Learning using Nonequilibrium Thermody- namics

    Jascha Sohl-Dickstein et al. “Deep Unsupervised Learning using Nonequilibrium Thermody- namics”. In:Proceedings of the 32nd International Conference on Machine Learning. Ed. by Francis Bach and David Blei. V ol. 37. Proceedings of Machine Learning Research. Lille, France: PMLR, July 2015, pp. 2256–2265.URL: https://proceedings.mlr.press/v37/ sohl-dickstei...

    2015

  7. [7]

    Flow Matching for Generative Modeling

    Yaron Lipman et al. “Flow Matching for Generative Modeling”. In:The Eleventh International Conference on Learning Representations. 2023.URL: https://openreview.net/forum? id=PqvMRDCJT9t(cit. on pp. 1, 3)

    2023

  8. [8]

    Qiang Liu.Rectified Flow: A Marginal Preserving Approach to Optimal Transport. 2022. arXiv: 2209.14577 [stat.ML] .URL: https://arxiv.org/abs/2209.14577 (cit. on pp. 1, 3)

    Pith/arXiv arXiv 2022

  9. [9]

    Genie: Generative Interactive Environments

    Jake Bruce et al. “Genie: Generative Interactive Environments”. In:Proceedings of the 41st International Conference on Machine Learning. Ed. by Ruslan Salakhutdinov et al. V ol. 235. Proceedings of Machine Learning Research. PMLR, 21–27 Jul 2024, pp. 4603–4623.URL: https://proceedings.mlr.press/v235/bruce24a.html(cit. on pp. 2, 9, 14)

    2024

  10. [10]

    Latent Action Pretraining from Videos

    Seonghyeon Ye et al. “Latent Action Pretraining from Videos”. In:The Thirteenth International Conference on Learning Representations. 2025.URL: https://openreview.net/forum? id=VYOe2eBQeh(cit. on pp. 2, 9)

    2025

  11. [11]

    Attention is All you Need

    Ashish Vaswani et al. “Attention is All you Need”. In:Advances in Neural Informa- tion Processing Systems. Ed. by I. Guyon et al. V ol. 30. Curran Associates, Inc., 2017. URL: https : / / proceedings . neurips . cc / paper _ files / paper / 2017 / file / 3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf(cit. on p. 4)

    2017

  12. [12]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    Jacob Devlin et al. “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding”. In:Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). Ed. by Jill Burstein, Christy Doran, and Thamar Solorio. Minneapolis, Mi...

    2019

  13. [13]

    Zhai and B

    William Peebles and Saining Xie. “ Scalable Diffusion Models with Transformers ”. In:2023 IEEE/CVF International Conference on Computer Vision (ICCV). Los Alamitos, CA, USA: IEEE Computer Society, Oct. 2023, pp. 4172–4182.DOI: 10.1109/ICCV51070.2023.00387. URL: https://doi.ieeecomputersociety.org/10.1109/ICCV51070.2023.00387 (cit. on p. 5)

    work page doi:10.1109/iccv51070.2023.00387 2023

  14. [14]

    Octo: An Open-Source Generalist Robot Policy

    Dibya Ghosh et al. “Octo: An Open-Source Generalist Robot Policy”. In:Proceedings of Robotics: Science and Systems. Delft, Netherlands, July 2024.DOI: 10.15607/RSS.2024. XX.090(cit. on pp. 6, 7, 9)

    work page doi:10.15607/rss.2024 2024

  15. [15]

    OpenVLA: An Open-Source Vision-Language-Action Model

    Moo Jin Kim et al. “OpenVLA: An Open-Source Vision-Language-Action Model”. In:8th Annual Conference on Robot Learning. 2024.URL: https://openreview.net/forum?id= ZMnD6QZAE6(cit. on pp. 6, 7). 10

    2024

  16. [16]

    Mustafa Shukor et al.SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics. 2025. arXiv: 2506 . 01844 [cs.LG].URL: https : / / arxiv . org / abs / 2506 . 01844(cit. on pp. 6, 7)

    2025

  17. [17]

    Learning to Act Anywhere with Task-centric Latent Actions

    Qingwen Bu et al. “Learning to Act Anywhere with Task-centric Latent Actions”. In:Proceed- ings of Robotics: Science and Systems. LosAngeles, CA, USA, June 2025.DOI: 10.15607/ RSS.2025.XXI.014(cit. on pp. 6, 7, 9, 14, 15)

    2025

  18. [18]

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware

    Tony Z. Zhao et al. “Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware”. In:Proceedings of Robotics: Science and Systems. Daegu, Republic of Korea, July 2023.DOI: 10.15607/RSS.2023.XIX.016(cit. on pp. 6, 7, 16)

    work page doi:10.15607/rss.2023.xix.016(cit 2023

  19. [19]

    Behavior Generation with Latent Actions

    Seungjae Lee et al. “Behavior Generation with Latent Actions”. In:Proceedings of the 41st International Conference on Machine Learning. Ed. by Ruslan Salakhutdinov et al. V ol. 235. Proceedings of Machine Learning Research. PMLR, 21–27 Jul 2024, pp. 26991–27008.URL: https://proceedings.mlr.press/v235/lee24y.html(cit. on pp. 6, 7)

    2024

  20. [20]

    QueST: Self-Supervised Skill Abstractions for Learning Continuous Control

    Atharva Mete et al. “QueST: Self-Supervised Skill Abstractions for Learning Continuous Control”. In:Advances in Neural Information Processing Systems. Ed. by A. Globerson et al. V ol. 37. Curran Associates, Inc., 2024, pp. 4062–4089.DOI: 10 . 52202 / 079017 - 0133.URL: https://proceedings.neurips.cc/paper_files/paper/2024/file/ 076c3e48fa502c660902105965f...

    2024

  21. [21]

    Length pass, checkpoint fail

    Philipp Wu et al. “GELLO: A General, Low-Cost, and Intuitive Teleoperation Framework for Robot Manipulators”. In:2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). 2024, pp. 12156–12163.DOI: 10.1109/IROS58592.2024.10801581 (cit. on p. 6)

    work page doi:10.1109/iros58592.2024.10801581 2024

  22. [22]

    Steering Your Diffusion Policy with Latent Space Reinforcement Learning

    Andrew Wagenmaker et al. “Steering Your Diffusion Policy with Latent Space Reinforcement Learning”. In:9th Annual Conference on Robot Learning. 2025.URL: https://openreview. net/forum?id=jU7AbGq3se(cit. on p. 9)

    2025

  23. [23]

    Parrot: Data-Driven Behavioral Priors for Reinforcement Learning

    Avi Singh et al. “Parrot: Data-Driven Behavioral Priors for Reinforcement Learning”. In: International Conference on Learning Representations. 2021.URL: https://openreview. net/forum?id=Ysuv-WOFeKR(cit. on p. 9)

    2021

  24. [24]

    Imitating Latent Policies from Observation

    Ashley Edwards et al. “Imitating Latent Policies from Observation”. In:Proceedings of the 36th International Conference on Machine Learning. Ed. by Kamalika Chaudhuri and Ruslan Salakhutdinov. V ol. 97. Proceedings of Machine Learning Research. PMLR, Sept. 2019, pp. 1755–1763.URL: https://proceedings.mlr.press/v97/edwards19a.html (cit. on p. 9)

    2019

  25. [25]

    Learning to Act without Actions

    Dominik Schmidt and Minqi Jiang. “Learning to Act without Actions”. In:International Conference on Learning Representations. Ed. by B. Kim et al. V ol. 2024. 2024, pp. 9379– 9395.URL: https : / / proceedings . iclr . cc / paper _ files / paper / 2024 / file / 27985d21f0b751b933d675930aa25022-Paper-Conference.pdf(cit. on p. 9)

    2024

  26. [26]

    Moto: Latent Motion Token as the Bridging Language for Learning Robot Manipulation from Videos

    Yi Chen et al. “Moto: Latent Motion Token as the Bridging Language for Learning Robot Manipulation from Videos”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Oct. 2025, pp. 19752–19763 (cit. on p. 9)

    2025

  27. [27]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. “Decoupled Weight Decay Regularization”. In:International Conference on Learning Representations. 2019.URL: https://openreview.net/forum? id=Bkg6RiCqY7(cit. on pp. 14, 17)

    2019

  28. [28]

    Mingxing Xu et al.Spatial-Temporal Transformer Networks for Traffic Flow Forecasting. 2021. arXiv: 2001.02908 [eess.SP] .URL: https://arxiv.org/abs/2001.02908 (cit. on p. 14)

    arXiv 2021

  29. [29]

    NSVQ: Noise Substitution in Vector Quantiza- tion for Machine Learning

    Mohammad Hassan Vali and Tom Bäckström. “NSVQ: Noise Substitution in Vector Quantiza- tion for Machine Learning”. In:IEEE Access10 (2022), pp. 13598–13610.DOI: 10.1109/ ACCESS.2022.3147670(cit. on p. 14)

    arXiv 2022

  30. [30]

    RT-1: Robotics Transformer for Real-World Control at Scale

    Anthony Brohan et al. “RT-1: Robotics Transformer for Real-World Control at Scale”. In: Proceedings of Robotics: Science and Systems. Daegu, Republic of Korea, July 2023.DOI: 10.15607/RSS.2023.XIX.025(cit. on pp. 14, 15)

    work page doi:10.15607/rss.2023.xix.025(cit 2023

  31. [31]

    DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset

    Alexander Khazatsky et al. “DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset”. In:Proceedings of Robotics: Science and Systems. Delft, Netherlands, July 2024.DOI: 10. 15607/RSS.2024.XX.120(cit. on p. 15). 11

    2024

  32. [32]

    2024 , url =

    Abby O’Neill et al. “Open X-Embodiment: Robotic Learning Datasets and RT-X Models : Open X-Embodiment Collaboration”. In:2024 IEEE International Conference on Robotics and Automation (ICRA). 2024, pp. 6892–6903.DOI: 10.1109/ICRA57147.2024.10611477 (cit. on p. 15)

    work page doi:10.1109/icra57147.2024.10611477 2024

  33. [33]

    Ego4d: Around the world in 3,000 hours of egocentric video

    Kristen Grauman et al. “Ego4d: Around the world in 3,000 hours of egocentric video”. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 18995–19012 (cit. on p. 15)

    2022

  34. [34]

    2024 (cit

    Remi Cadene et al.LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch.https://github.com/huggingface/lerobot. 2024 (cit. on p. 16)

    2024

  35. [35]

    Narasimhan, and Shunyu Yao

    Benjamin Warner et al. “Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference”. In:Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Ed. by Wanxiang Che et al. Vienna, Austria: Association for Computational Linguis...

    work page doi:10.18653/v1/2025 2025

  36. [36]

    Deep Residual Learning for Image Recognition

    Kaiming He et al. “Deep Residual Learning for Image Recognition”. In:Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 2016 (cit. on p. 16)

    2016

  37. [37]

    Redmon, S

    Jia Deng et al. “ImageNet: A large-scale hierarchical image database”. In:2009 IEEE Confer- ence on Computer Vision and Pattern Recognition. 2009, pp. 248–255.DOI: 10.1109/CVPR. 2009.5206848(cit. on p. 16). 12 A Qualitative analysis figures We provide Figures 3 and 4, which allow us to better understand our method qualitatively. Figure 3 shows that latent...

    work page doi:10.1109/cvpr 2009