arxiv: 2605.11383 · v1 · submitted 2026-05-12 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

HamBR: Active Decision Boundary Restoration Based on Hamiltonian Dynamics for Learning with Noisy Labels

Ningkang Peng , Jingyang Mao , Qianfeng Yu , Xiaoqian Peng , Peirong Ma , Yanhui Gu

Authors on Pith no claims yet

Pith reviewed 2026-05-13 02:07 UTC · model grok-4.3

classification 💻 cs.CV

keywords noisy label learningdecision boundary restorationHamiltonian Monte Carlosemi-supervised learningimage classificationout-of-distribution detectionenergy-based modelingboundary collapse

0 comments

The pith

A Hamiltonian dynamics method restores collapsed decision boundaries in noisy-label learning by synthesizing virtual outliers that push samples toward class centers.

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

Noisy labels cause decision boundaries in neural networks to collapse in feature space, making it hard for models to separate clean but hard samples from mislabeled ones. The paper proposes HamBR, which uses spherical Hamiltonian Monte Carlo to actively explore these collapsed regions and generate virtual outlier samples. These outliers create energy barriers that repel samples away from the boundaries toward their class centers. This restores sharpness to the boundaries and improves overall accuracy in noisy label settings. The method can be plugged into existing frameworks and also boosts out-of-distribution detection.

Core claim

The paper claims that actively restoring decision boundaries by synthesizing high-quality virtual outliers using Spherical Hamiltonian Monte Carlo probing of inter-class ambiguous regions, and imposing energy barriers through energy-based modeling, forces samples to move toward class centers and restores discriminative sharpness for better noise-robust learning in DNNs.

What carries the argument

Spherical Hamiltonian Monte Carlo mechanism for probing inter-class ambiguous regions and synthesizing virtual outliers that establish energy barriers through energy-based modeling.

If this is right

Significantly enhances accuracy for hard boundary samples in noisy label scenarios.
Achieves state-of-the-art performance when integrated into semi-supervised noisy label learning frameworks on CIFAR-10, CIFAR-100, and real-world noise datasets.
Provides superior convergence efficiency and robustness.
Improves the model's ability to detect out-of-distribution samples.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The boundary restoration idea might extend to other settings where feature overlap occurs, such as long-tailed class distributions.
Combining the energy barrier approach with different sampling methods could yield faster or more scalable variants.
The technique's effect on model calibration and uncertainty estimates could be measured in follow-up experiments.

Load-bearing premise

That Spherical HMC probing of inter-class ambiguous regions will reliably synthesize high-quality virtual outliers whose imposed energy barriers restore discriminative sharpness without introducing new artifacts or harming clean-sample performance.

What would settle it

A test on a synthetic dataset with controlled label noise where feature dispersion within classes is measured before and after applying the method; if dispersion does not decrease and hard-sample accuracy does not rise, the restoration claim fails.

Figures

Figures reproduced from arXiv: 2605.11383 by Jingyang Mao, Ningkang Peng, Peirong Ma, Qianfeng Yu, Xiaoqian Peng, Yanhui Gu.

**Figure 2.** Figure 2: Overview of the HamBR Framework.The framework comprises a classification branch for sample partitioning and a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: (a) Ablation study of the HamBR module and verification of its versatility within the DivideMix framework. (b) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

read the original abstract

In large-scale visual recognition and data mining tasks, the presence of noisy labels severely undermines the generalization capability of deep neural networks (DNNs). Prevalent sample selection methods rely primarily on training loss or prediction confidence for passive screening. However, within a feature space degraded by noise, decision boundaries undergo systematic boundary collapse. This phenomenon hinders the ability of the model to distinguish between hard clean samples and noisy samples at the decision margins, thereby creating a significant performance bottleneck. This study is the first to emphasize the pivotal importance of active boundary restoration for noise-robust learning. We propose HamBR, a novel paradigm based on Hamiltonian dynamics. The core approach leverages the Spherical Hamiltonian Monte Carlo (Spherical HMC) mechanism to actively probe inter-class ambiguous regions within the representation space and synthesize high-quality virtual outliers. By imposing explicit repulsion constraints via energy-based modeling, these synthesized samples establish robust energy barriers at the decision boundaries. This mechanism forces real samples to move from dispersed overlapping regions toward their respective class centers, thereby restoring the discriminative sharpness of the decision boundaries. HamBR demonstrates exceptional versatility and can be integrated as a plug-and-play defense module into existing semi-supervised noisy label learning frameworks. Empirical evaluations show that the proposed paradigm significantly enhances the discriminative accuracy of hard boundary samples, achieving state-of-the-art (SOTA) performance on CIFAR-10/100 and real-world noise benchmarks. Furthermore, it exhibits superior convergence efficiency and reliable robustness, while improving significantly the capability of the model for Out-of-Distribution (OOD) detection.

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.

Desk Editor's Note private letter to a colleague

HamBR applies Spherical HMC to generate virtual outliers that impose energy repulsion and restore collapsed decision boundaries under noisy labels, but the abstract supplies no derivations, ablations, or numbers to check whether the mechanism actually works.

read the letter

The main thing to know is that this paper frames noisy-label problems as boundary collapse in feature space and proposes active restoration via Hamiltonian dynamics. It uses Spherical HMC to probe ambiguous inter-class regions, synthesize virtual outliers, and apply energy-based repulsion so that real samples shift toward class centers. The method is presented as a plug-and-play module for existing semi-supervised noisy-label pipelines, with claimed gains on CIFAR-10/100, real-world noise sets, and OOD detection. That combination of dynamics-based probing and explicit boundary barriers is new relative to standard loss or confidence selection approaches, and the problem framing itself is useful even if the execution needs checking. The paper does a decent job identifying why passive screening alone hits a wall once noise has already blurred the margins between hard clean samples and mislabeled ones. The OOD improvement is a reasonable side claim if the energy barriers generalize beyond the training distribution. The soft spots are straightforward. The abstract gives no equations for the Hamiltonian, the repulsion term, or how Spherical HMC is adapted to the learned feature space. There are no ablations on the number or placement of virtual outliers, no error bars, and no checks that the synthesized points avoid landing inside clean manifolds or pushing hard clean samples into worse overlap. The central assumption—that HMC trajectories will reliably produce high-quality outliers whose barriers sharpen boundaries without new artifacts—remains untested in the provided text. If the energy function is misspecified or the feature space is too multimodal, the repulsion could degrade performance on clean data instead of helping. This work is for researchers already working on label-noise robustness in vision, especially those who want to move past sample selection toward explicit boundary fixes. A reader who follows DivideMix-style methods or energy-based models could extract the core idea and test it. I would send it to peer review because the idea is distinct enough that referees can examine the full implementation, run the ablations, and see whether the SOTA numbers hold up with proper controls.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes HamBR, a plug-and-play module for noisy-label learning that employs Spherical Hamiltonian Monte Carlo (Spherical HMC) to actively probe inter-class ambiguous regions in feature space, synthesize virtual outliers, and impose energy-based repulsion constraints. These constraints are claimed to restore collapsed decision boundaries by driving real samples toward class centers, yielding SOTA accuracy on CIFAR-10/100 and real-world noise benchmarks while also improving convergence and OOD detection.

Significance. If the empirical claims and the boundary-restoration mechanism hold, the work would be significant as the first explicit emphasis on active (rather than passive) boundary restoration in noisy-label settings. The plug-and-play design could be integrated into existing semi-supervised frameworks, and the reported gains in hard-boundary accuracy and OOD performance would be practically useful. No machine-checked proofs or parameter-free derivations are presented.

major comments (3)

Abstract: the claim of achieving 'state-of-the-art (SOTA) performance on CIFAR-10/100 and real-world noise benchmarks' is unsupported by any numerical results, tables, ablation studies, or error bars, rendering the central performance claim unevaluable from the provided text.
Method description (Spherical HMC and energy-based modeling): the repulsion constraints and virtual-outlier synthesis are described only at a high level with no equations, pseudocode, or analysis of trajectory stability, mode collapse, or risk that synthesized points lie inside clean manifolds; this step is load-bearing for the claim that boundaries are restored without degrading clean-sample performance.
Abstract (boundary restoration claim): the assertion that imposed energy barriers 'force real samples to move from dispersed overlapping regions toward their respective class centers' lacks any derivation, guarantee against new artifacts, or check on clean hard-sample displacement, which directly underpins the noise-robustness argument.

minor comments (2)

Abstract: the phrase 'virtual outliers' is introduced without a precise definition or distinction from standard outlier synthesis techniques.
Abstract: the statement that HamBR 'exhibits superior convergence efficiency' is not accompanied by any training-curve or iteration-count evidence.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We have carefully reviewed each major comment and provide point-by-point responses below, outlining how we will strengthen the presentation of our results, method, and theoretical claims through targeted revisions.

read point-by-point responses

Referee: Abstract: the claim of achieving 'state-of-the-art (SOTA) performance on CIFAR-10/100 and real-world noise benchmarks' is unsupported by any numerical results, tables, ablation studies, or error bars, rendering the central performance claim unevaluable from the provided text.

Authors: We acknowledge that the abstract summarizes the SOTA claim without embedding specific numbers or tables, which is standard for brevity. The full manuscript contains the supporting evidence in Section 4, with quantitative comparisons, ablation studies, and error bars across multiple runs. To address the concern directly, we will revise the abstract to explicitly reference the experimental section (e.g., 'achieving state-of-the-art performance as shown in our experiments on CIFAR-10/100 and real-world benchmarks'). This makes the claim traceable without altering the abstract's length constraints. revision: yes
Referee: Method description (Spherical HMC and energy-based modeling): the repulsion constraints and virtual-outlier synthesis are described only at a high level with no equations, pseudocode, or analysis of trajectory stability, mode collapse, or risk that synthesized points lie inside clean manifolds; this step is load-bearing for the claim that boundaries are restored without degrading clean-sample performance.

Authors: The manuscript presents the core equations for Spherical HMC dynamics and the energy-based repulsion in Section 3, along with pseudocode in the supplementary material. However, we agree that additional analysis would strengthen the load-bearing step. In the revision, we will expand Section 3 with explicit equations for the repulsion term, include the full pseudocode in the main text, and add a dedicated paragraph analyzing trajectory stability under the spherical constraint, the mitigation of mode collapse via temperature annealing, and empirical verification (via feature-space visualizations and distance metrics) that synthesized outliers remain outside clean manifolds. These additions will clarify how boundary restoration occurs without harming clean-sample performance. revision: yes
Referee: Abstract (boundary restoration claim): the assertion that imposed energy barriers 'force real samples to move from dispersed overlapping regions toward their respective class centers' lacks any derivation, guarantee against new artifacts, or check on clean hard-sample displacement, which directly underpins the noise-robustness argument.

Authors: We will augment the method section with a step-by-step derivation showing how the gradient of the energy-based repulsion term induces the desired movement of samples toward class centers. While the work is empirical and does not offer a formal guarantee against all possible artifacts, the revised manuscript will include explicit checks on clean hard-sample displacement (reporting accuracy on verified clean subsets before and after applying HamBR) to demonstrate stability. These additions will be placed in Section 3.3 and cross-referenced in the abstract to better support the noise-robustness argument. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces HamBR as a novel paradigm leveraging Spherical HMC to probe ambiguous regions and synthesize virtual outliers for active boundary restoration in noisy label learning. The abstract and description present this as an original mechanism with explicit repulsion constraints via energy-based modeling, without any equations, fitted parameters renamed as predictions, or load-bearing self-citations that reduce the central claim to its own inputs. The approach is framed as a plug-and-play module integrable with existing frameworks, with SOTA claims resting on empirical benchmarks rather than self-referential definitions or ansatzes smuggled via prior work. This keeps the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only view limits visibility; the approach assumes Hamiltonian dynamics can be meaningfully applied to representation spaces and that synthesized outliers will act as effective barriers.

axioms (1)

domain assumption Hamiltonian dynamics govern movement and energy in the feature representation space
Invoked to justify Spherical HMC probing of ambiguous regions

invented entities (1)

virtual outliers no independent evidence
purpose: Establish robust energy barriers at decision boundaries via repulsion constraints
Synthesized samples used to force real samples toward class centers

pith-pipeline@v0.9.0 · 5594 in / 1286 out tokens · 47462 ms · 2026-05-13T02:07:58.745119+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By imposing explicit repulsion constraints via energy-based modeling, these synthesized samples establish robust energy barriers at the decision boundaries.
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Spherical Hamiltonian Monte Carlo (Spherical HMC) mechanism to actively probe inter-class ambiguous regions

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages · 2 internal anchors

[1]

Devansh Arpit, Stanisław Jastrzębski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. 2017. A closer look at memorization in deep networks. In International conference on machine learning. PMLR, 233–242

work page 2017
[2]

Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jiatong Li, Yinian Mao, Gang Niu, and Tongliang Liu. 2021. Understanding and improving early stopping for learning with noisy labels.Advances in Neural Information Processing Systems34 (2021), 24392–24403

work page 2021
[3]

David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. 2019. Mixmatch: A holistic approach to semi-supervised learning.Advances in neural information processing systems32 (2019)

work page 2019
[4]

Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101–mining discriminative components with random forests. InEuropean conference on com- puter vision. Springer, 446–461

work page 2014
[5]

Charles Bouveyron and Stéphane Girard. 2009. Robust supervised classifica- tion with mixture models: Learning from data with uncertain labels.Pattern Recognition42, 11 (2009), 2649–2658

work page 2009
[6]

Gaoyu Cao, Zhanquan Sun, Chaoli Wang, Hongquan Geng, Hongliang Fu, Zhong Yin, and Minlan Pan. 2024. RASNet: Renal automatic segmentation using an im- proved U-Net with multi-scale perception and attention unit.Pattern Recognition 150 (2024), 110336

work page 2024
[7]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInter- national conference on machine learning. PmLR, 1597–1607

work page 2020
[8]

Filipe R Cordeiro, Ragav Sachdeva, Vasileios Belagiannis, Ian Reid, and Gustavo Carneiro. 2023. Longremix: Robust learning with high confidence samples in a noisy label environment.Pattern recognition133 (2023), 109013

work page 2023
[9]

Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4690–4699

work page 2019
[10]

Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. 2022. Vos: Learning what you don’t know by virtual outlier synthesis.arXiv preprint arXiv:2202.01197 (2022)

work page arXiv 2022
[11]

Jacob Goldberger and Ehud Ben-Reuven. 2017. Training deep neural-networks using a noise adaptation layer. InInternational conference on learning representa- tions

work page 2017
[12]

Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets.Advances in neural information processing systems27 (2014)

work page 2014
[13]

Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels.Advances in neural information processing systems31 (2018)

work page 2018
[14]

Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Mo- mentum contrast for unsupervised visual representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9729–9738

work page 2020
[15]

Dan Hendrycks and Kevin Gimpel. 2016. A baseline for detecting misclas- sified and out-of-distribution examples in neural networks.arXiv preprint arXiv:1610.02136(2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[16]

Nazmul Karim, Mamshad Nayeem Rizve, Nazanin Rahnavard, Ajmal Mian, and Mubarak Shah. 2022. Unicon: Combating label noise through uniform selection and contrastive learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9676–9686

work page 2022
[17]

Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning.Advances in neural information processing systems33 (2020), 18661–18673

work page 2020
[18]

Shiwei Lan, Bo Zhou, and Babak Shahbaba. 2014. Spherical Hamiltonian Monte Carlo for constrained target distributions. InInternational Conference on Machine Learning. PMLR, 629–637

work page 2014
[19]

Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems31 (2018)

work page 2018
[20]

Hengzhuang Li and Teng Zhang. 2025. Outlier synthesis via hamiltonian monte carlo for out-of-distribution detection.arXiv preprint arXiv:2501.16718(2025)

work page arXiv 2025
[21]

Junnan Li, Richard Socher, and Steven CH Hoi. 2020. Dividemix: Learning with noisy labels as semi-supervised learning.arXiv preprint arXiv:2002.07394(2020)

work page arXiv 2020
[22]

Shikun Li, Xiaobo Xia, Shiming Ge, and Tongliang Liu. 2022. Selective-supervised contrastive learning with noisy labels. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 316–325

work page 2022
[23]

Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda

work page
[24]

Early-learning regularization prevents memorization of noisy labels.Ad- vances in neural information processing systems33 (2020), 20331–20342

work page 2020
[25]

Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. 2020. Energy-based Out- of-distribution Detection. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 33. 21464–21475

work page 2020
[26]

Radford M Neal et al. 2011. MCMC using Hamiltonian dynamics.Handbook of markov chain monte carlo2, 11 (2011), 2

work page 2011
[27]

Curtis Northcutt, Lu Jiang, and Isaac Chuang. 2021. Confident learning: Esti- mating uncertainty in dataset labels.Journal of Artificial Intelligence Research70 (2021), 1373–1411

work page 2021
[28]

Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. 2017. Making deep neural networks robust to label noise: A loss correction approach. InProceedings of the IEEE conference on computer vision and pattern recognition. 1944–1952

work page 2017
[29]

Baoye Song, Shihao Zhao, Luyao Dang, Haoguang Wang, and Lin Xu. 2025. A survey on learning from data with label noise via deep neural networks.Systems Science & Control Engineering13, 1 (2025), 2488120

work page 2025
[30]

Hwanjun Song, Minseok Kim, and Jae-Gil Lee. 2019. Selfie: Refurbishing unclean samples for robust deep learning. InInternational conference on machine learning. PMLR, 5907–5915

work page 2019
[31]

Haoliang Sun, Chenhui Guo, Qi Wei, Zhongyi Han, and Yilong Yin. 2022. Learning to rectify for robust learning with noisy labels.Pattern Recognition124 (2022), 108467

work page 2022
[32]

Haobo Wang, Ruixuan Xiao, Yixuan Li, Lei Feng, Gang Niu, Gang Chen, and Junbo Zhao. 2022. PiCO: Contrastive Label Disambiguation for Partial Label Learning.ICLR1, 2 (2022), 5

work page 2022
[33]

Xiaobo Xia, Tongliang Liu, Bo Han, Nannan Wang, Mingming Gong, Haifeng Liu, Gang Niu, Dacheng Tao, and Masashi Sugiyama. 2020. Part-dependent label noise: Towards instance-dependent label noise.Advances in neural information processing systems33 (2020), 7597–7610

work page 2020
[34]

Ruixuan Xiao, Yiwen Dong, Haobo Wang, Lei Feng, Runze Wu, Gang Chen, and Junbo Zhao. 2022. Promix: Combating label noise via maximizing clean sample utility.arXiv preprint arXiv:2207.10276(2022)

work page arXiv 2022
[35]

Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama

work page
[36]

In International conference on machine learning

How does disagreement help generalization against label corruption?. In International conference on machine learning. PMLR, 7164–7173

work page
[37]

Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2017. mixup: Beyond empirical risk minimization.arXiv preprint arXiv:1710.09412 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[38]

Qian Zhang, Yi Zhu, Filipe R Cordeiro, and Qiu Chen. 2025. PSSCL: A progressive sample selection framework with contrastive loss designed for noisy labels. Pattern Recognition161 (2025), 111284

work page 2025
[39]

Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen

work page
[40]

arXiv preprint arXiv:2103.07756(2021)

Learning with feature-dependent label noise: A progressive approach. arXiv preprint arXiv:2103.07756(2021)

work page arXiv 2021
[41]

Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. 2019. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1237–1246

work page 2019
[42]

Chenyue Zhou, Mingxuan Wang, Yanbiao Ma, Chenxu Wu, Wanyi Chen, Zhe Qian, Xinyu Liu, Yiwei Zhang, Junhao Wang, Hengbo Xu, et al . 2025. From perception to cognition: A survey of vision-language interactive reasoning in multimodal large language models.arXiv preprint arXiv:2509.25373(2025)

work page arXiv 2025