Pith. sign in

REVIEW 3 major objections 5 minor 26 references

EXAONE Path 2.0: Pathology Foundation Model with End-to-End Supervision

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a pathology foundation model trained under direct slide-level supervision on only 37,195 whole-slide images reaches state-of-the-art average AUROC (0.784) across 10 biomarker prediction tasks, outperforming larger…

desk verdict A plausible training recipe with a missing data-hygiene statement that undermines the headline claim of SOTA data efficiency. read the letter →

arxiv 2507.06639 v2 pith:UEDDM7UH submitted 2025-07-09 cs.CV cs.AIcs.LG

Myeongjang Pyeon , Janghyeon Lee , Minsoo Lee , Juseung Yun , Hwanil Choi , Jonghyun Kim , Jiwon Kim , Yi Hu
show 2 more authors
Jongseong Jang Soonyoung Lee
This is my paper
classification cs.CVcs.AIcs.LG
keywords pathologyfoundationmodelwhole-slideimagesslide-levelsupervisionbiomarkerpredictionhierarchicalvisiontransformercurriculumlearningmulti-taskdataefficiency
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

EXAONE Path 2.0 is a pathology foundation model whose central claim is that patch-level representations should be learned under direct slide-level supervision rather than only through self-supervised patch pretraining. Using 37,195 whole-slide images, the model jointly trains a three-stage hierarchical vision transformer on cancer subtyping, tissue classification, and molecular biomarker labels, then adapts to downstream tasks with a lightweight multiple-instance-learning aggregator. The paper reports the best average AUROC (0.784) among seven foundation models over ten biomarker prediction tasks spanning mutation status and microsatellite instability, while using fewer training slides than several competitors. If this result is correct, clinically relevant features for biomarker prediction can be learned with substantially less training data than SSL-based approaches require, and the standard SSL-then-aggregate recipe is not the only route to a strong pathology foundation model.

What carries the argument

The load-bearing mechanism is a three-stage Hierarchical Image Pyramid Transformer (HIPT) trained end to end: a first-stage ViT encodes 256 by 256 patches, a second-stage ViT aggregates patch embeddings into region representations, and a third-stage ViT integrates region representations into a full-slide embedding, with slide-level cross-entropy loss flowing through all stages. A two-stage curriculum starts with DINO self-supervised losses at patch and region scales and then turns on supervised loss at 4096 by 4096 region resolution, while activation checkpointing and CPU offloading make gigapixel training tractable. For downstream adaptation, the model exits the hierarchy early, pooling frozen first-stage patch features with CLAM, a clustering-constrained attention multiple-instance-learning aggregator, instead of fine-tuning the full network.

What would settle it

Compare the slide identifiers or hashed specimen IDs of the 37,195 pretraining WSIs with the test slides of the 10 benchmark tasks; any overlap between pretraining and a benchmark test set would invalidate the reported AUROC advantage. Alternatively, rerun the same evaluation on a fresh cohort of WSIs from a hospital that contributed no training slides; if the 0.784 average does not survive, the claim of data-efficient generalization collapses.

Watch

Extended reading notes

Core claim

The central discovery the authors are trying to establish is that end-to-end slide-level supervision during pretraining makes a pathology foundation model both more accurate and more data-efficient than patch-level SSL encoders. On 10 slide-level benchmarks spanning lung, colorectal, breast, renal, and colon cancer, EXAONE Path 2.0 obtains an average AUROC of 0.784, compared with 0.765 for PRISM, 0.755 for UNI2-h, 0.748 for TITAN, 0.731 for EXAONE Path 1.0, 0.707 for Prov-GigaPath, and 0.696 for CHIEF. The authors attribute this result to slide-level labels propagating gradients through all hierarchical stages, so that the first-stage patch features themselves carry clinically relevant biomarker information. Their claim is that this supervision, not model scale or pretraining slide count, is what drives the gain.

Load-bearing premise

The load-bearing premise is that none of the slides used in the 10 benchmark test sets appeared among the 37,195 pretraining slides, because the pretraining tasks are the same label types and the cohorts overlap; the paper never states this separation.

Editorial extensions

If this is right

  • If the central claim holds, pathology foundation models can be pretrained on tens of thousands of labeled slides instead of hundreds of thousands of unlabeled whole-slide images, lowering data and compute barriers.
  • Slide-level labels encoding molecular outcomes produce first-stage patch features that transfer directly to mutation and microsatellite-instability prediction, so downstream tasks should need fewer task-specific labeled slides.
  • The reported average AUROC of 0.784, about 0.02 to 0.09 above the SSL-trained baselines, indicates that direct supervision is a competitive alternative to pure self-supervised pretraining.
  • Because downstream adaptation reuses the same frozen patch features with a standard MIL aggregator, existing CLAM-based pipelines can adopt the model without architectural changes.

Reading between the lines

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

  • The paper does not report an AUROC-versus-training-slide-count curve, so the specific claim that 37k slides are sufficient rests on a single comparison point; retraining on 5k, 15k, and 25k slides would show whether the gains are monotone and where saturation appears.
  • The evaluation assumes benchmark test slides are disjoint from the 37,195 pretraining slides, but the paper never states this; checking slide or specimen identifiers for overlap between pretraining and the CPTAC, KOR, and USA test sets would settle whether the reported scores reflect generalization or memorization.
  • One could test the mechanism directly by ablating the multi-task slide labels one at a time; if removing tissue classification degrades mutation prediction, the shared-representation story is supported, and if not, the gains may come mainly from the largest task.
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

3 major / 5 minor

Summary. The paper proposes EXAONE Path 2.0, a hierarchical Vision Transformer for pathology whole-slide images trained end-to-end with direct slide-level supervision. Using 37,195 WSIs with 144,450 image-label pairs across 16 tasks (cancer subtyping, tissue classification, and molecular biomarker prediction), the authors report an average AUROC of 0.784 over 10 biomarker prediction benchmarks, claiming state-of-the-art performance compared to TITAN, PRISM, CHIEF, Prov-GigaPath, UNI2-h, and EXAONE Path 1.0. The method employs curriculum learning and memory management to make gigapixel-scale training feasible. The evaluation uses frozen backbones with linear or CLAM heads and reports averages over four seeds.

Significance. If the results are valid, the paper makes a useful contribution by showing that direct slide-level supervision can be computationally practical for gigapixel pathology images and can achieve strong performance with fewer slides than SSL-only foundation models. The multi-seed evaluation and the comparison with seven baselines are positive elements. However, the central claim of state-of-the-art performance depends on an explicit data-hygiene guarantee and adequate statistical reporting, neither of which is currently present. The paper does not provide code, and the appended license agreement does not substitute for reproducibility details. The potential value is real, but the evidence as presented is insufficient to support the headline conclusion.

major comments (3)
  1. [§2.2, §3.1, §3.4] The manuscript never states that the 37,195 pretraining WSIs (§3.1) are disjoint from the benchmark test slides (§3.4). This is load-bearing because §2.2 lists pretraining tasks of cancer subtyping across 33 types, tissue classification, and molecular biomarker prediction (including MSI, mutation status, and hormone receptor), which are exactly the label types used in the 10 evaluation tasks. Concretely, CRC-MSI in §3.4.1 is trained on KOR-CRC and tested on a 'held-out portion of the same dataset'; if KOR-CRC slides were among the pretraining WSIs, the test slides may have been seen with MSI labels during pretraining. Likewise, the CPTAC-BRCA/CCRCC/COAD datasets are public and could have been included in the 33-cancer-type or mutation pretraining tasks. The phrase 'predefined training/test split' in §3.3 appears to refer only to the downstream CLAM fine-tuning split, not to pretraining overlap. Without an explicit identifier-level disjointness guarantee, the reported average AUROC advantage over PRISM (0.784 vs 0.765) and UNI2-h (0.755) may reflect in-distribution memorization rather than genuine generalization. The authors must provide a slide-level overlap analysis or exclude all benchmark test slides from pretraining.
  2. [§3.3, Table 1, §3.4.2] Although §3.3 states that each benchmark was evaluated over four independent seeds, Table 1 reports only point estimates with no standard deviations or confidence intervals. This is especially problematic because many test sets are very small: BRCA-TP53 has 22 test slides, BRCA-PIK3CA has 21, RCC-PBRM1 has 52, RCC-BAP1 has 50, COAD-KRAS has 19, and COAD-TP53 has 18. The average AUROC differences among top models are on the order of 0.02–0.03, which may not be statistically distinguishable at these sample sizes. The paper should report seed-wise standard deviations and, ideally, a paired significance test across the 10 tasks (e.g., Wilcoxon signed-rank) to support the 'state-of-the-art average performance' claim.
  3. [§2.2, §3.4, Figure 1b] The evaluation tasks are in-distribution with respect to the pretraining objectives: the model was pretrained to predict MSI, mutation status, and cancer subtype, and the benchmarks evaluate the same label types. This does not automatically invalidate the results, but it weakens the 'generalization' and 'foundation model' interpretation and makes the data-efficiency comparison in Figure 1b difficult to interpret. The 37k pretraining WSIs are paired with 144,450 slide-level labels (§3.1), whereas the SSL baselines use unlabeled slide corpora; comparing only the number of WSIs conflates supervised annotation effort with raw data quantity. The authors should clarify what 'data-efficient' means in this context and, if possible, add held-out tasks that were not part of the pretraining label set to demonstrate transfer beyond the training task families.
minor comments (5)
  1. [§3.1] The paper states that 37,195 WSIs generate 144,450 image-label pairs across 16 training tasks, but §2.2 describes only three broad task families; please provide a complete list of the 16 tasks with label sources and cohort sizes for reproducibility.
  2. [Table 1] Table 1 should include standard deviations in addition to the reported AUROC point estimates, at least as a supplementary table, to support the multi-seed evaluation described in §3.3.
  3. [§3.5] The statement that EXAONE Path 2.0 'remained within the top three across nearly all tasks' should be quantified: list the tasks where it is not in the top three and provide exact rank positions.
  4. [Appendix] The appended EXAONE Path model license agreement is not part of the technical content and should be moved to supplementary material or removed to avoid confusing readers about the paper's scientific scope.
  5. [§3.4.2] For the CPTAC benchmarks, please include the exact data collection versions and any preprocessing steps (e.g., tissue segmentation, staining normalization, patch extraction protocol) to improve reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Benchmark test splits are only defined for fine-tuning, not for the 37,195-WSI supervised pretraining, whose objectives include the exact biomarker labels being 'predicted'; the SOTA claim is therefore partially forced by construction unless pretraining/test disjointness is established.

  1. fitted input called prediction [§2.2 Multi-Task Learning; §3.1 Training Data; §3.3 Evaluation Protocols; §3.4.1 CRC-MSI benchmark]
    "These WSIs generate 144,450 image-label pairs across 16 training tasks ... including 'molecular biomarker prediction including pan-cancer and cancer-specific mutation status, microsatellite instability, and hormone receptor subtyping' ... 'Each benchmark task was evaluated on a predefined training/test split' ... 'CRC-MSI. ... tested on a held-out portion of the same dataset (658:209)'."

    The only hold-out guarantee in the paper is the 'predefined training/test split' of §3.3, which is defined for downstream CLAM fine-tuning and says nothing about the 37,195-WSI pretraining corpus. The pretraining objectives in §2.2 explicitly include the same label types used in the 10 benchmarks (MSI, KRAS/TP53/PIK3CA/PBRM1/BAP1 mutation status, cancer subtyping), and the benchmark cohorts are private KOR/USA hospital collections plus public CPTAC, which are not stated to be excluded from pretraining. In particular, CRC-MSI is tested on a held-out portion of the same KOR-CRC dataset used for CLAM training, but the paper never establishes that this held-out portion was excluded from slide-level supervised pretraining.

full rationale

Most of the paper is methodologically transparent: no uniqueness theorem is imported from the authors' prior work, no ansatz is smuggled in via citation, and the self-citation to EXAONE Path 1.0 is only a baseline, not load-bearing. The circularity concern is concentrated in the evaluation protocol. The supervised pretraining objectives include the exact biomarker labels used in the ten benchmarks, and the paper's only 'held-out' language applies to the downstream CLAM split, not to the 37,195-WSI pretraining corpus. Because the benchmark test sets (KOR-CRC held-out, CPTAC) are not shown to be disjoint from the pretraining data, the reported SOTA AUROCs cannot be read as pure generalization; if any test slides carry pretraining labels, the result is an in-distribution fit. This matches the 'fitted input called prediction' pattern: the evaluation is presented as a prediction even though the pretraining stage may already have fitted the very labels and slides being evaluated. An explicit identifier-level disjointness statement would remove the circularity and drop the score to 0-2; as written, the central claim is unverifiable and partially circular.

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

The central claim rests on data hygiene (disjoint test splits), statistical inference from tiny test sets, and a fair comparison protocol. The paper does not supply evidence for these premises. It introduces no new entities, free parameters, or mathematical axioms beyond standard deep learning assumptions.

assumptions (3)
  • domain assumption Benchmark test sets are disjoint from the 37k pretraining WSIs
    This is not stated in the paper, yet the central claim of generalization depends on it. Label overlap between pretraining tasks and benchmarks makes this assumption non-trivial.
  • domain assumption AUROC differences of about 0.02 on tiny test sets are statistically meaningful
    The paper interprets small AUROC gaps as ranking evidence without confidence intervals or hypothesis tests.
  • domain assumption The frozen-backbone plus CLAM protocol is a fair comparison across all patch-level models
    All patch-level models use the same aggregator, but the pretraining objectives differ; the comparison may favor the model whose pretraining included the target labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EXAONE Path 2.0: Pathology Foundation Model with End-to-End Supervision." pith.science (2026). https://pith.science/paper/UEDDM7UH

@misc{pith2026250706639,
  author       = {Pith},
  title        = {Pith review of: EXAONE Path 2.0: Pathology Foundation Model with End-to-End Supervision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UEDDM7UH}},
  note         = {Machine review of arXiv:2507.06639}
}
read the original abstract

In digital pathology, whole-slide images (WSIs) are often difficult to handle due to their gigapixel scale, so most approaches train patch encoders via self-supervised learning (SSL) and then aggregate the patch-level embeddings via multiple instance learning (MIL) or slide encoders for downstream tasks. However, patch-level SSL may overlook complex domain-specific features that are essential for biomarker prediction, such as mutation status and molecular characteristics, as SSL methods rely only on basic augmentations selected for natural image domains on small patch-level area. Moreover, SSL methods remain less data efficient than fully supervised approaches, requiring extensive computational resources and datasets to achieve competitive performance. To address these limitations, we present EXAONE Path 2.0, a pathology foundation model that learns patch-level representations under direct slide-level supervision. Using only 37k WSIs for training, EXAONE Path 2.0 achieves state-of-the-art average performance across 10 biomarker prediction tasks, demonstrating remarkable data efficiency.

Figures

Figures reproduced from arXiv: 2507.06639 by the authors.

Figure 1
Figure 1. Performance comparison of models based on the number of parameters and the number [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. End-to-end hierarchical learning in EXAONE Path 2.0. Slide-level supervised signals [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Comparison of AUROC scores across 10 slide-level benchmarks and their averages. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 17 canonical work pages

  1. [1]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650–9660, 2021

    2021

  2. [2]

    Scaling vision transformers to gigapixel images via hierarchical self-supervised learning

    Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144–16155, 2022

    2022

  3. [3]

    Towards a general-purpose foundation model for computational pathology

    Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Bowen Chen, Andrew Zhang, Daniel Shao, Andrew H Song, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 2024

    2024

  4. [4]

    Multimodal whole slide foundation model for pathology

    Tong Ding, Sophia J Wagner, Andrew H Song, Richard J Chen, Ming Y Lu, Andrew Zhang, Anurag J Vaidya, Guillaume Jaume, Muhammad Shaban, Ahrong Kim, et al. Multimodal whole slide foundation model for pathology. arXiv preprint arXiv:2411.19666, 2024

    arXiv 2024

  5. [5]

    The cptac data portal: a resource for cancer proteomics research

    Nathan J Edwards, Mauricio Oberti, Ratna R Thangudu, Shuang Cai, Peter B McGarvey, Shine Jacob, Subha Madhavan, and Karen A Ketchum. The cptac data portal: a resource for cancer proteomics research. Journal of proteome research, 14(6):2707–2713, 2015. 6

    work page 2015

  6. [6]

    Shallow-deep networks: Understanding and mitigating network overthinking

    Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitras. Shallow-deep networks: Understanding and mitigating network overthinking. In International conference on machine learning , pages 3301–3310. PMLR, 2019

    work page 2019

  7. [7]

    Where Should I Spend My FLOPS? Efficiency Evaluations of Visual Pre-training Methods

    Skanda Koppula, Yazhe Li, Evan Shelhamer, Andrew Jaegle, Nikhil Parthasarathy, Relja Arand- jelovic, João Carreira, and Olivier Hénaff. Where should i spend my flops? efficiency evaluations of visual pre-training methods. arXiv preprint arXiv:2209.15589, 2022

    work page Pith review arXiv 2022

  8. [8]

    Data-efficient and weakly supervised computational pathology on whole-slide images

    Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature biomedical engineering, 5(6):555–570, 2021

    work page 2021

Show all 26 references
  1. [9]

    The clinical proteomic tumor analysis consortium clear cell renal cell carcinoma collection (cptac-ccrcc) (version 13) [data set]

    National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC). The clinical proteomic tumor analysis consortium clear cell renal cell carcinoma collection (cptac-ccrcc) (version 13) [data set]. https://doi.org/10.7937/k9/tcia.2018.oblamn27, 2018

    doi

  2. [10]

    The clinical proteomic tumor analysis consortium breast invasive carcinoma collection (cptac-brca) (version 1) [data set]

    National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC). The clinical proteomic tumor analysis consortium breast invasive carcinoma collection (cptac-brca) (version 1) [data set]. https://doi.org/10.7937/TCIA.CAEM-YS80, 2020

    doi

  3. [11]

    The clinical proteomic tumor analysis consortium colon adenocarcinoma collection (cptac-coad) (version 1) [data set]

    National Cancer Institute Clinical Proteomic Tumor Analysis Consortium (CPTAC). The clinical proteomic tumor analysis consortium colon adenocarcinoma collection (cptac-coad) (version 1) [data set]. https://doi.org/10.7937/TCIA.YZWQ-ZZ63, 2020

    doi

  4. [12]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  5. [13]

    Prism: A multi-modal generative foundation model for slide-level histopathology.arXiv preprint arXiv:2405.10254, 2024

    George Shaikovski, Adam Casson, Kristen Severson, Eric Zimmermann, Yi Kan Wang, Jeremy D Kunz, Juan A Retamero, Gerard Oakley, David Klimstra, Christopher Kanan, et al. Prism: A multi-modal generative foundation model for slide-level histopathology.arXiv preprint arXiv:2405.10...

    2024 arXiv

  6. [14]

    Solving inefficiency of self-supervised representation learning

    Guangrun Wang, Keze Wang, Guangcong Wang, Philip HS Torr, and Liang Lin. Solving inefficiency of self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9505–9515, 2021

    2021

  7. [15]

    A pathology foundation model for cancer diagnosis and prognosis prediction

    Xiyue Wang, Junhan Zhao, Eliana Marostica, Wei Yuan, Jietian Jin, Jiayu Zhang, Ruijiang Li, Hongping Tang, Kanran Wang, Yu Li, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, 634(8035):970–978, 2024

    2024

  8. [16]

    Wright, Ari Robicsek, Brian Piening, Carlo Bifulco, Sheng Wang, and Hoifung Poon

    Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, Yanbo Xu, Mu Wei, Wenhui Wang, Shuming Ma, Furu Wei, Jianwei Yang, Chunyuan Li, Jianfeng Gao, Jaylen Rosemon, Tucker Bower, Soohee Lee, Roshant...

    2024

  9. [17]

    Agreement

    Juseung Yun, Yi Hu, Jinhyung Kim, Jongseong Jang, and Soonyoung Lee. Exaonepath 1.0 patch-level foundation model for pathology. arXiv preprint arXiv:2408.00380, 2024. 7 EXAONEPath AI Model License Agreement 1.0 - NC This License Agreement (“Agreement”) is entered into between ...

    2024 arXiv

  10. [18]

    Definitions 1.1 Model The artificial intelligence model provided by Licensor, which includes any software, algorithms, machine learning models, or related components supplied by Licensor. This definition extends to encompass all updates, enhancements, improvements, bug fixes, ...

  11. [19]

    EXAONEPath

    License Grant 2.1 Grant of License Subject to the terms and conditions outlined in this Agreement, the Licensor hereby grants the Licensee a limited, non-exclusive, non-transferable, worldwide, and revocable license to: a. Access, download, install, and use the Model solely fo...

  12. [20]

    Any commercial exploitation of the Model or its derivatives requires a separate commercial license agreement with the Licensor

    Restrictions 3.1 Commercial Use The Licensee is expressly prohibited from using the Model, Derivatives, or Output for any commercial purposes, including but not limited to, developing or deploying products, services, or applications that generate revenue, whether directly or i...

  13. [21]

    The Licensee acknowledges that this Agreement does not transfer any ownership rights to the Licensee

    Ownership 4.1 Intellectual Property All rights, title, and interest in and to the Model, including any modifications, Derivatives, and associated documentation, are and shall remain the exclusive property of the Licensor. The Licensee acknowledges that this Agreement does not ...

  14. [22]

    As-Is” Basis The Model, Derivatives, and Output are provided on an “as-is

    No Warranty 5.1 “As-Is” Basis The Model, Derivatives, and Output are provided on an “as-is” and “as-available” basis, without any warranties or representations of any kind, whether express, implied, or statutory. The Licensor disclaims all warranties, including but not limited...

  15. [23]

    Limitation of Liability 6.1 No Liability for Damages To the fullest extent permitted by applicable law, in no event shall the Licensor be liable for any special, incidental, indirect, consequential, exemplary, or punitive damages, including but not limited to, damages for loss...

  16. [24]

    Termination shall be effective immediately upon notice

    Termination 7.1 Termination by Licensor The Licensor reserves the right to terminate this Agreement and revoke the Licensee’s rights to use the Model at any time, with or without cause, and without prior notice if the Licensee breaches any of the terms or conditions of this Ag...

  17. [25]

    Governing Law 8.1 Governing Law This Agreement shall be governed by and construed in accordance with the laws of the Republic of Korea, without regard to its conflict of laws principles. 8.2 Arbitration Any disputes, controversies, or claims arising out of or relating to this ...

  18. [26]

    Any modifications will be effective upon posting the updated Agreement on the Licensor’s website or through other means of communication

    Alterations 9.1 Modifications The Licensor reserves the right to modify or amend this Agreement at any time, in its sole discretion. Any modifications will be effective upon posting the updated Agreement on the Licensor’s website or through other means of communication. The Li...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.