List of Papers Browse by Subject Areas Author List
Abstract
Pathology Visual Grounding aims to precisely localize pathological tissue regions in histopathology images and output bounding boxes based on professional medical language descriptions. However, in histopathology images, background textures are highly similar, while target regions are small, low-contrast, and blurred at the boundaries. In addition, medical descriptions can be difficult to understand accurately, making it challenging to associate the described textual location with a unique target region. Existing pathology visual grounding methods make insufficient use of fine-grained visual cues and rely more on coarse language-vision alignment, which can easily cause misalignment and drift in difficult samples, such as small targets, occlusion, or blurred boundaries. To address these issues, we propose PCRP (Progressive Coarse-to-Fine), a progressive framework for pathology visual grounding that improves grounding accuracy by strengthening reliable visual cues and performing robust boundary refinement. PCRP consists of four modules: Visual Branch, Text Branch, VWF, and C2F-BAR. This design reduces misalignment and drift in samples with small targets and blurred boundaries, thereby improving robustness and grounding quality. PCRP achieves state-of-the-art performance on RefPath and significantly outperforms PKNet by 11.68% Acc and 9.27% mIoU, validating its effectiveness. The source code is available at https://github.com/wuhengliangliang/PCRP.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4084_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wuhengliangliang/PCRP
Link to the Dataset(s)
N/A
BibTex
@InProceedings{PenLia_PCRP_MICCAI2026,
author = { Peng, Liang AND Li, Chenxiao AND Dong, Shaohua AND Tan, Bohan AND Zhang, Zhipeng AND Dong, Xingping},
title = { { PCRP: Progressive Coarse-to-Fine Refinement for Pathology Grounding } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper studies pathology visual grounding, aiming to localize pathological regions in histopathology images from medical language descriptions. The authors propose PCRP, a progressive coarse-to-fine framework with two main components: Visual-Word Fusion (VWF) for cross-modal alignment and C2F-BAR for boundary-aware box refinement using SAM-based mask priors. The method achieves strong results on the RefPath benchmark, especially on the more challenging 20× setting.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Important task setting. Pathology visual grounding is a meaningful and relatively underexplored problem in computational pathology, especially because precise region localization from clinical language has practical value.
Strong empirical performance. The reported results on RefPath are impressive, with clear gains over prior methods, particularly on low-magnification images where grounding is more difficult.
Reasonable overall framework. The coarse-to-fine design is intuitive, and the use of visual cue enhancement plus mask-based refinement is well aligned with the challenges of small targets and blurry boundaries in pathology images.
Clear presentation. The paper is generally easy to follow, and the ablation study helps explain the contribution of each module.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Limited methodological novelty. In my view, the main weakness is that the proposed method is largely a combination of existing ideas rather than a clearly new technical contribution. Cross-attention-based vision-language fusion, gated residual updates, coarse-to-fine refinement, curriculum-style training, and the use of SAM-generated priors are all familiar design patterns. While the integration is reasonable, the paper does not sufficiently demonstrate what is fundamentally new beyond adapting these components to pathology grounding.
Heavy reliance on external priors weakens the novelty claim. A substantial part of the performance gain seems to come from using SAM3 as a teacher to provide mask priors for refinement. This is practically useful, but it makes the improvement feel more like a strong engineering pipeline built on top of an external foundation model, rather than a novel grounding method in itself. The paper should better clarify how much of the gain comes from the proposed architecture versus the external segmentation prior.
Experimental validation is still somewhat limited. Although the gains on RefPath are large, evaluation is restricted to a single benchmark. This makes it difficult to judge generalizability. In addition, the paper does not report variance across runs, statistical significance, or deeper analysis of failure cases. Given the surprisingly large improvement over previous work, stronger validation would be important.
Some claims are stronger than the evidence provided. The paper repeatedly emphasizes robustness on difficult cases such as tiny targets, blurred boundaries, and low magnification, but the supporting analysis is still relatively limited. More targeted breakdowns or controlled experiments would make these claims more convincing.
Overall, I think this is a well-motivated and practically useful paper, and the empirical results are strong. However, for me the technical novelty is not sufficient, and the current experimental evidence is not yet strong enough to fully support the scale of the claimed advance. Therefore, my current recommendation is weak reject.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important pathology grounding task and reports strong empirical results on RefPath. However, the main components—cross-modal fusion, coarse-to-fine refinement, curriculum-style transition, and SAM-based prior guidance—are largely incremental combinations of existing ideas. Given the limited methodological novelty and the lack of broader validation beyond a single benchmark, I lean weak reject.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes PCRP, a pathology visual grounding framework for localizing pathology regions from referring expressions in histopathology images. The method combines: (i) Visual-Word Fusion (VWF) to inject image-conditioned evidence into token-level textual representations, and (ii) C2F-BAR, a two-stage coarse-to-fine box regression scheme that uses SAM3-derived mask priors with reliability weighting and progressive prompt transition to refine boundaries under box-only supervision. On the RefPath benchmark, the method reports substantial gains over prior pathology visual grounding approaches, particularly on the more difficult 20× split.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.A reasonable and clear structure for the pathology task 2.Good picture, the drawing enables easy following of the methodology 3.Good research motivation for solving the tiny pathology target 4.Good Result 5.C2F-BAR is an easy example that demonstrates how to utilize a SAM-like prior-knowledge model more effectively.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The weakness is obvious in this work. I list the major weaknesses from the top-priority to the bottom.
1.The refinement design is methodologically questionable because Stage-2 is driven by SAM3 masks obtained from box prompts, and during early training those prompts are ground-truth boxes. This introduces a strong privileged localization prior into the refinement process, making it unclear whether the reported gains come from better pathology grounding or simply from external mask-based correction under privileged supervision.
This means that the authors claimed that this work relies on vision-language interaction to find the object. However, the import of prior mask localization for the SAM3 actually gains from the object area leakage from ground truth.(mask area -> SAM3 -> prior knowledge)
I couldn’t argue that PCRP achieves good performance in this task, at least in their validated dataset. BUT I am afraid that the major improvement is from the attention area leakage via SAM3. 2.When excluding the SAM3 and corresponding enhancement module C2F-BAR. The VWF is essentially text-to-vision cross-attention with gated residual fusion. The novelty is moderate.
3.In real pathology workflows, the key challenge is often to detect tiny or easily overlooked abnormal regions that may not have been pre-localized by a human or an upstream model. However, the main gain of PCRP comes from Stage-2 refinement driven by SAM3 masks obtained from a box prompt. This design mainly improves localization around an already available candidate region, rather than demonstrating the ability to discover missed tiny targets. Therefore, the method seems better characterized as a candidate-refinement framework than a clinically meaningful solution for reducing oversight of subtle pathology findings.
4.All EXPERIMENT is only validated on a SINGLE DATASET.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
I have two additional comments for authors, and I hope they might help authors.
1.Rethinking the motivation for this framework, the prior knowledge should be carefully designed and set, rather than directly importing the GroundTruth. The real clinical scenarios are impossible to provide the GT mask. When you are trying to utilize the GT, the improvement and contribution already shift from the real clinical impact to the leaderboard Illusion.
2.If authors really wanna dive into the real clinical contribution, you might try pseudo-label by encoder’s attention, or the grid and point prompt design for SAM3.It could enable the segmentation to really work if it neglects the inference latency of the system.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is relevant to MIC, and the reported gains on RefPath are substantial, which makes it a borderline submission rather than an outright reject. However, I am unconvinced that the paper has isolated its true contribution. A central part of the performance improvement appears to come from incorporating SAM3-derived mask priors, yet the experiments do not adequately distinguish gains due to the proposed model design from gains due to using a strong external segmentation prior. In addition, the novelty is moderate, the ablation is not diagnostic enough, and the evaluation is limited to a single dataset without external validation or robustness analysis.
Overall, the paper is promising, but in its current form, it falls slightly below the MICCAI acceptance threshold.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Authors respond to reviewers appropriately, and some concerns have already been solved.
Review #3
- Please describe the contribution of the paper
This paper proposes PCRP, a progressive coarse-to-fine framework for pathology visual grounding. The method tackles two main issues in this task: semantic drift in language-image alignment and inaccurate localization under weak or blurry boundaries. To address these, the authors introduce Visual-Word Fusion (VWF) to inject image-conditioned visual evidence into word-level text features, and C2F-BAR, a two-stage regression framework that refines coarse predictions using confidence-weighted mask priors from a frozen SAM3 teacher. On RefPath, the method achieves substantial gains over previous approaches, with particularly strong improvements on the more challenging 20× subset.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
(i) The paper studies a meaningful and relatively underexplored problem in computational pathology, namely visual grounding from pathology-specific referring expressions. (ii) The quantitative results are strong. The qualitative examples are also helpful.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(i) The main limitation is that the evaluation is restricted to RefPath only. The results on this benchmark are strong, but it is still hard to tell how well the method would generalize beyond this dataset or to other pathology grounding settings. (ii) The method depends on a frozen SAM3 teacher for mask priors, but the paper does not really analyze how sensitive the final performance is to the quality of these priors. This is particularly relevant because the Stage-2 refinement may be vulnerable when the Stage-1 box is already off-target. (iii) The paper would be stronger with a clearer discussion of computational overhead. Since the method introduces a two-stage refinement pipeline and relies on external mask generation, some reporting of runtime or inference cost would help readers judge the practical trade-off. (iv) While the ablations are useful, the analysis remains somewhat high level. It would be helpful to see more failure case analysis, especially for cases where the SAM3 prior is noisy or the referring expression is ambiguous. (v) There are a few presentation issues that should be cleaned up in the final version. For instance, the reference list appears to contain a duplicated citation for PathVG, and a few parts of the method section could be written more clearly.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
(i) References [18] and [19] seem to be the same PathVG paper, with the same authors, title, venue, and page numbers. The reference list and in-text citations should be cleaned up accordingly. (ii) The paper would be stronger with more discussion of generalization beyond RefPath, since the current evaluation is limited to a single benchmark. (iii) A brief analysis of runtime or computational overhead would also be helpful, given the two-stage design and the additional use of SAM3-based priors. (iv) In the Stage-2 refinement part, the refined prediction should correspond to b1, and the loss function later also explicitly supervises b1 and b0.However, Eq. (8) is written as b0 = VLT(…), which is likely a notation error and should probably be b1 (v) The dependence on the quality of SAM3-generated mask priors deserves more analysis, especially in cases where the Stage-1 proposal is inaccurate.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I am leaning Weak Accept. The main reason is that the paper makes a fairly clear methodological contribution to pathology visual grounding and backs it up with strong empirical gains on the main benchmark. In particular, the improvement on the challenging 20× subset is substantial, which supports the authors’ claim that the method is especially helpful in difficult low-magnification settings. However, the evaluation is limited to one dataset and the practical implications of using SAM3-based mask priors are not fully analyzed. Overall, I think the paper is above the bar, but not by a large margin.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers and the AC for their constructive comments. We address the major concerns below.
Q1: Gains attributed to SAM3 priors. [AC,R1,R2,R3]
SAM3-derived priors provide useful boundary information, but PCRP’s improvement cannot be simply attributed to SAM3.The first row of Table 2 is already a SAM3-assisted PKNet-style baseline. On 20×, SAM3 prior substantially improves PKNet, increasing Acc by +13.64; however, on 40×, its standalone gain is limited, increasing Acc by only +1.92.In contrast, on top of this SAM3-assisted baseline, full PCRP further improves Acc by +2.92 and +2.64 on 20× and 40×. This shows that SAM3 alone cannot explain the final performance; VWF, CBP, and PPT provide additional gains under different magnifications, especially on the 40× baseline.
Q2: Potential GT leakage through SAM3.[R2]
Sorry for the confusion. We would like to clarify that GT is NOT leaked, because the GT will be used ONLY in the training process. During inference, GT is NOT used for prediction. We will clarify this in revision.
Q3: Problem-driven formulation beyond known components. [R1,R2]
We understand the reviewer’s concern. While some components are inspired by existing techniques, our novelty lies in the coarse-to-fine pathology-grounding framework. Our method combines coarse localization for tiny/subtle abnormal regions with SAM3-guided refinement, enabling a coarse-to-fine solution specifically designed for pathology grounding rather than a simple module combination. The strong empirical results further show that this design is effective, achieving clear improvements over prior methods. We will better clarify the framework-level novelty and its performance advantages in the revision.
Q4: RefPath-only evaluation. [AC,R1,R2,R3]
We understand the importance of external dataset validation. However, pathology visual grounding is a new task, and to our best knowledge, RefPath is currently the only public pathology visual grounding dataset. For this reason, we conduct evaluation of our method on RefPath. It is worth noting, RefPath contains two test subsets with different magnifications, including 40× and 20×, and uses various metrics for evaluation. Our PCRP achieves state-of-the-art results on both subsets, showing its efficacy and robustness in handling images of different resolution in real scenarios. We will clarify this in revision. Thanks!
Q5: Reliability, failure cases, and overhead. [AC,R1,R3]
PCRP doesn’t directly use SAM3 masks as final prediction, but converts them into boundary cues and injects them into stage-2 for grounding. CBP suppresses noisy priors caused by inaccurate prompts, so the model does not unconditionally rely on SAM3 prior. Severe stage-1 deviation, noisy SAM3 prior, or ambiguous expression are possible failure cases. Fig. 2 in the submitted paper has qualitatively shown that stage-2 can alleviate part of the grounding drift. Regarding computational overhead, two-stage refinement and external mask generation will bring additional cost. We will provide discussion about the accuracy-efficiency trade-off and failure modes in the revised version.
Q6: Candidate refinement vs. missed-target discovery. [R2]
Thanks for this insightful comment. We agree that discovering tiny or overlooked pathological regions is important in clinical practice. Our framework is designed for tiny objects and subtle abnormal regions. Stage-1 performs coarse localization to capture a potential abnormal region, and stage-2 further refines localization and boundary. Thus, PCRP is not merely candidate refinement, but progressively handles subtle pathology targets through coarse-to-fine localization. We will clarify this motivation.
Q7: Reproducibility. [R2]
We will release our full source code and model, and provide more implementation details in the revised paper.
Q8: Presentation errors. [R3]
Thanks for pointing this out. We will double-check and carefully polish the paper in revision. Again, thanks!
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper presents a clear and empirically strong pathology visual grounding method, but its overall standing remains borderline because the strongest gains are closely tied to the SAM3-assisted refinement stage and the validation is limited to a single benchmark. The strengths of the submission are the relevance of the task, the clear coarse-to-fine formulation, and the large performance improvements reported on RefPath, including especially strong gains on the more difficult 20x subset. The paper also provides a basic ablation showing that VWF, CBP, and PPT each contribute to the final result, and the method is generally presented in a coherent and understandable way. The decisive concerns are threefold. First, the method relies materially on SAM3-derived mask priors, and the manuscript itself indicates that much of the gain comes from this refinement stage; this raises a legitimate question about how much of the contribution is due to the proposed framework design versus the external prior. Second, the empirical validation is restricted to RefPath, so the current evidence for generalization remains limited. Third, the analysis of sensitivity to prior quality, failure modes, and computational overhead is limited. At the same time, after checking the manuscript, some reviewer criticisms are only partially supported. In particular, the concern that the method directly uses ground-truth masks is incorrect: the paper uses SAM3-derived masks and states that ground-truth boxes are used only in an early training warm-up before transitioning to predicted boxes, with inference aligned to the latter. Accordingly, I assign limited weight to the stronger leakage claim. Similarly, the criticism that the method is clinically misaligned because it performs refinement is not persuasive for a visual grounding paper, where a coarse-to-fine localization design is consistent with the task being studied.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The main concerns remain the moderate methodological novelty and the dependence on an external segmentation prior. Several components, including cross-modal fusion, gated residual updates, coarse-to-fine refinement, curriculum-style prompt transition, and SAM-based priors, are individually familiar. The contribution is therefore best viewed as a task-specific integration and refinement framework rather than a fundamentally new grounding paradigm. In addition, the strongest gains appear closely related to the SAM3-assisted refinement stage, and the submitted paper does not fully isolate the effect of the proposed architecture from the benefit of using a strong external prior. The evaluation is also limited to RefPath, without external validation, variance estimates, statistical significance analysis, detailed sensitivity to prior quality, or runtime/overhead analysis. The rebuttal clarified an important point: the method does not use ground-truth masks at inference time, and the use of ground-truth boxes is restricted to an early training warm-up before transitioning to predicted boxes. This reduces the weight of the strongest leakage concern raised by R2.The rebuttal also argued that the improvements cannot be attributed solely to SAM3, since the ablation baseline is described as already SAM3-assisted and the full framework adds further gains through VWF, CBP, and PPT. This clarification is useful, although it does not completely remove the concern that a large part of the total improvement over prior work is enabled by the external SAM3 prior. The response to the single-dataset concern is also reasonable to some extent, given the limited availability of public pathology visual grounding benchmarks, but the generalization evidence remains limited. In reconciling the reviews, I gave substantial weight to the consistent agreement that the task is relevant, the method is clearly presented, and the quantitative results are strong. I also considered R1’s continued concern about incremental novelty and limited validation, R2’s initial concern about privileged priors but subsequent post-rebuttal acceptance, and R3’s weak accept position based on the strong benchmark performance despite practical and generalization limitations. Overall, the paper is above the acceptance threshold, but only marginally.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes a two-stage, coarse-to-fine, progressive framework for pathology visual grounding on the RefPath benchmark. The reviewers acknowledged the clarity of the formulation, the strong gains on the harder magnification subset, and the per-module ablation. The decisive concerns were reliance on SAM3 mask priors and the attribution of gains; potential ground-truth leakage through SAM3 prompted by training-time boxes; evaluation restricted to a single benchmark; and limited analysis of sensitivity to prior quality and failure modes. The rebuttal addresses the leakage concern by clarifying that ground-truth prompts are used only during training and never at inference, and by disentangling the SAM3-prior contribution from that of the proposed modules, as reported in the ablation. The single-benchmark concern is partly mitigated by the fact that RefPath is currently the only public pathology visual grounding dataset. The most consequential reviewer concerns are substantively addressed, and the practical contribution is meaningful. The camera-ready must include the failure-case analysis and the computational-overhead reporting promised in the rebuttal.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The scores after rebuttal are mixed. The contribution remains heavily tied to SAM3-derived refinement priors, making it difficult to isolate the proposed framework’s novelty from the external segmentation prior. Evaluation is limited to a single benchmark, with insufficient analysis of prior quality sensitivity, failure modes, computational overhead, and generalization. The method appears more like an engineering pipeline than a sufficiently novel grounding approach.
