List of Papers Browse by Subject Areas Author List
Abstract
Surgical vision-language foundation models typically adopt educational materials, such as surgical lecture videos, to transfer surgical knowledge encoded in language into visual representations. These knowledge are multi-dimensional and hierarchical: fine-grained action cues appear in narration, mid-level key steps are summarized in subsection headings, and global procedural context, such as patient history and surgical strategy, is described in abstract texts. Prior work largely collapses these heterogeneous signals into a single flat embedding space, implicitly assuming independence across hierarchy levels. However, this is suboptimal because it ignores cross-level semantic containment, e.g., actions belong to steps, steps compose phases, weakens long-range dependency modeling.
To this end, we propose a hyperbolic surgical video-language pre-training framework that explicitly preserves the hierarchical structure by mitigating structural false negatives induced by procedural context and enforcing semantic consistency between parent phases and their constituent child steps.
Extensive experiments on multiple surgical benchmarks show consistent gains in zero- and few-shot phase recognition across procedures and institutions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5295_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HuYao_HyperVLP_MICCAI2026,
author = { Hu, Yaojun AND Yuan, Kun AND Navab, Nassir AND Ying, Haochao AND Wu, Jian AND Padoy, Nicolas},
title = { { HyperVLP: Enhancing Hierarchical Surgical Video-Language Pre-training in Hyperbolic Space } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16892},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces HyperVLP, a surgical video-language pre-training framework that models the hierarchical structure of surgical workflows (actions to steps to phases) in hyperbolic space. By employing geometry-aware contrastive learning and cone-based entailment loss, the method explicitly models parent-child relations rather than treating video-text pairs as flat clips. The framework is evaluated on multiple surgical datasets, showing improved representation quality for cross-procedure and cross-center generalization.
- 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.
- Clear Motivation: The method logically maps the natural hierarchy of surgical workflows to the mathematical properties of hyperbolic space.
- Strong Empirical Results: The zero-shot and few-shot linear probing results are encouraging across multiple benchmarks (Cholec80, AutoLaparo, StrasBypass70, BernBypass70). The gains are especially clear on AutoLaparo (e. g. , zero-shot accuracy improving from 26.5 / 23.6 to 42.9 / 32.9 over PeskaVLP).
- Appropriate Loss Formulation: The ablation study effectively shows that combining geometry-aware contrastive learning and cone-based entailment learning contributes positively to performance
- 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.
- Unclear Euclidean vs. Hyperbolic Gains: The paper lacks a direct Euclidean counterpart equipped with the same hierarchical losses to confirm that gains stem fundamentally from the hyperbolic manifold rather than the added hierarchy-aware supervision alone.
- Fairness of Comparison: It is not fully transparent whether the compared methods (SurgVLP, HecVL, PeskaVLP) use the exact same backbone, pre-training data volume, and text supervision.
- Corpus & Hierarchy Metrics Details: The construction of the SVL corpus lacks detail on how reliable parent-child relations and ASR transcripts were ensured. Additionally, the qualitative radial-distance analysis (Figure 2) lacks quantitative hierarchy-preservation metrics to support it.
- Parameter Stability & Uncertainty: The paper does not discuss the stability or selection process for key parameters (curvature c, cone parameter K, η, τ). It also omits uncertainty estimates (confidence intervals/standard deviations) for the downstream results.
- 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
- Euclidean Baseline: Discuss or provide a small-scale comparison against a Euclidean baseline using the exact same hierarchical constraints to isolate the effect of hyperbolic geometry.
- Baseline Fairness & Uncertainty: Please clarify if prior methods were evaluated under strictly identical conditions (backbone, data). Adding standard deviations for the linear probing results would also strengthen the claims.
- Parameter Stability Discussion: Since a full sensitivity analysis is computationally prohibitive for pre-training, please add a brief discussion on how key parameters (c, K, η, τ) were selected and share observations on optimization stability.
- Dataset & Metric Clarifications: Briefly clarify how noisy hierarchy labels were filtered from the SVL dataset. Consider adding a quantitative metric to support the qualitative claims made in Figure 2.
- Notation Fixes: Clarify the V_c notation and improve the readability of the adaptive weight and entailment angle formulas.
- 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 presents a conceptually sound and experimentally effective approach to hierarchical surgical video-language pre-training. The empirical results, particularly the large gains on AutoLaparo and Stras70, are very promising. However, the manuscript needs to clarify baseline fairness, dataset construction, and the Euclidean versus hyperbolic distinction. Furthermore, fixing notation ambiguities and discussing parameter stability are necessary. If these clarifications and discussions are provided during the rebuttal, this would make for a very solid contribution to MICCAI.
- Reviewer confidence
Very confident (4)
- [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.
The authors have offered clear explanations regarding the fairness of the Euclidean baselines (HecVL, PeskaVLP) and have carefully defined the structural sets S(i) using temporal metadata rather than semantic assumptions. These responses help clarify the theoretical points I initially raised. While step-level evaluations (such as on MultiBypass140), which could further support claims of hierarchical consistency, are not yet included and are expected in the final version due to space constraints, I am keeping my score at Weak Reject. I have chosen this score because, while I recognize the conceptual soundness, I believe that the forthcoming empirical validations will be particularly valuable.
Review #2
- Please describe the contribution of the paper
This paper proposes a hierarchy-aware vision-language pre-training framework for surgical videos. Unlike standard flat clip–caption alignment approaches, this work explicitly models the hierarchical levels of the surgical data by introducing hyperbolic embeddings to model semantic relations across multiple levels (clip, phase segment, and video). The approach is evaluated primarily on phase recognition tasks under both zero-shot and few-shot settings.
- 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.
- Novel formulation for surgical vision language pre-training. The proposed approach moves beyond flat euclidean space and uses hyperbolic geometry for hierarchy aware learning to encode both similarity and containment to align with the nature of surgical workflow data. -The method demonstrates consistent improvements on multiple surgical phase recognition benchmarks under both zero-shot and few-shot settings, indicating that the learned representations transfer effectively across datasets.
- Modeling surgical workflows at multiple semantic levels is highly relevant for various downstream CAI applications.
- 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.
-
Gap between hierarchical modeling objective and evaluations. The method is designed to model hierarchical structure, but evaluation is limited to flat phase classification. As a result, it makes it difficult to assess the proposed hierarchical modeling. Further evaluations like cross modal retrieval are needed.
-
The paper does not sufficiently disentangle the effects of modified contrastive objectives, and hyperbolic geometry, making it difficult to assess which components are most responsible for the performance gains.
-
- 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
- If evaluation is only phase recognition, does hyperbolic geometry actually help? Because they are just flat labels. Can’t a strong Euclidean model with good supervision match performance?
- Including tasks that test the hierarchy nature can be significant to demonstrate the value of the proposed approach. May be retrieval tasks like Parent-to-Child retrieval? or Table 2 in PeskaVLP?
- Multibypass 140 also has step labels. Presenting these results could not add another evaluation, but also help validate hierarchical-consistency (e.g. whether step predicted is consistent with the phase)
- An ablation with comparing Euclidean vs hyperbolic embeddings with identical hierarchical supervision could be valuable in supporting the claims.
- I found it difficult to follow how the partition of candidates are being selected. While the idea of structural false negatives is clear, the definitions weren’t clear enough for me. Partly because the examples used to explain were not consistent sometimes with phases, and procedures. I think picking one parent-child example to explain this would make it easy for the readers to follow.
- What are Wv and Wt in 2.2 first paragraph? I couldn’t find any reference to this. Is this the projection layer being learned in stage 2?
- Minor: In 2.2 first paragraph, Vc is repeated twice the second term should be Tc, and following I think you meant to say “child-level features v and t”
- 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 presents a well-motivated and technically sound approach with strong empirical results and clear relevance to the surgical community. However, the central claim—improved hierarchical representation via hyperbolic geometry—is not fully validated by the current evaluation, which primarily focuses on flat phase recognition tasks. Additionally, the contribution of individual components is not clearly disentangled. While promising, the paper would benefit from further evaluations linking the proposed formulation to its claimed advantages.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
The authors rebuttal clarifies several concerns regarding the hierarchical modeling objective and evaluation strategy. While evaluations directly probing hierarchical containment (e.g parent child retrieval shown in Peska) could further strengthen interpretation, the methodology is well motivated and supported by benchmarks. Some uncertainty remains regarding the extent to which improvement arises specifically from hyperbolic modeling vs broader structured contrastive formulations (e.e.g handling of stuctural false negatives and contextual proximity which may not be unique to hyperbolic space). However, I believe the work presents a technically motivated approach with sufficient empirical support and potential interest to the surgical AI community.
Review #3
- Please describe the contribution of the paper
The main contribution of this manuscript lies in its conceptual reframing of surgical video-language pre-training, shifting the paradigm from flat Euclidean embeddings to a geometry-aware hierarchical framework in hyperbolic space. Specifically, the authors make three methodological contributions:
1.Reframing Hierarchical Workflows in Hyperbolic Space Unlike conventional methods that collapse heterogeneous, multi-dimensional signals into a single flat Euclidean embedding space , the proposed HyperVLP projects visual and textual representations into a shared Lorentz hyperbolic manifold. This approach successfully preserves the intrinsic tree structure of surgical workflows—where fine-grained actions belong to steps, and steps compose phases. This geometric shift is highly valuable for supporting low-distortion embeddings of semantic containment and improving long-range dependency modeling.
2.Mitigating Structural False Negatives via Geometry-Aware Contrastive Learning To intelligently align representations, the authors introduce a novel geometry-aware hyperbolic contrastive strategy. Rather than treating all non-matching video-text pairs as absolute negatives , this mechanism identifies structural false negatives induced by contextual proximity and hierarchical entailment. This allows the model to selectively down-weight semantically related samples within the same procedural context , offering a more nuanced fusion strategy that prevents the artificial separation of clinically correlated nodes and preserves the overall workflow topology.
3.Enforcing Semantic Consistency via Hyperbolic Entailment Cones The framework dynamically enforces multi-level semantic containment using a cone-based hyperbolic entailment objective. This design not only imposes geometric containment between parent and child embeddings across inter-modal relations but also elegantly accommodates intra-modal structural constraints.
- 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.
The manuscript presents a well-motivated and methodologically rigorous approach to surgical video-language pre-training. The major strengths of this work include:
1.Hyperbolic Space for Hierarchical Workflows: The paper elegantly reframes surgical video-language representation learning by shifting from a standard flat Euclidean embedding space to a geometry-aware hyperbolic paradigm. Instead of collapsing heterogeneous, multi-dimensional signals into a single flat space, HyperVLP maps representations into a shared Lorentz hyperbolic manifold. This is a highly original formulation because hyperbolic space, with its negative curvature and exponential capacity, naturally supports the low-distortion embedding of the inherent tree-structured hierarchy in surgical workflows (where fine-grained actions compose steps, and steps compose phases).
2.Mitigating Structural False Negatives: The authors provide a strong critical analysis of standard contrastive learning objectives, which assume all non-matching video-text pairs are independent negatives. They astutely identify that this introduces “structural false negatives” by pushing apart clinically related or contextually adjacent video segments, thereby distorting the workflow topology. To resolve this, the proposed geometry-aware hyperbolic contrastive loss intelligently down-weights semantically related samples within the same procedural context. This offers a much more nuanced alignment strategy that prevents over-separation of semantically correlated clips and preserves the hierarchical topology.
- 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.
Q1: Lack of Theoretical Guarantees for Hierarchy Preservation via Heuristic Reweighting. The authors propose mitigating “structural false negatives” to preserve the workflow hierarchy using an adaptive weighting mechanism, $\alpha_{ik}$, which relies on similarity normalization. However, this approach is fundamentally a heuristic continuous relaxation rather than a rigorous structural constraint. Because it relies on relative distances in the embedding space to infer structural correlations, there is no theoretical guarantee that this soft-weighting mechanism actually recovers or preserves a discrete tree structure. In real-world scenarios with noisy data (e.g., ASR transcription errors) or suboptimal initialization, this mechanism might mistakenly attenuate the penalty for true negative samples, potentially leading to a catastrophic “hierarchy collapse.” Given that this approach appears to be empirical “loss engineering” rather than strict structural modeling, how can the authors theoretically prove or empirically guarantee that this mechanism consistently reconstructs the valid surgical tree topology without being derailed by spurious similarities?
Q2: Ambiguity in Temporal Boundaries and the Definition of Structurally Related Samples $\mathcal{S}(i)$. The authors construct their hierarchical framework based on specific temporal units: Phases, Steps, and Actions. However, the methodology for establishing the exact boundaries of these continuous video segments is not explicitly detailed. This ambiguity becomes particularly problematic in the formulation of the geometry-aware contrastive loss, where the authors introduce the set $\mathcal{S}(i)$, loosely defined as “structurally related samples from the same procedure (including contextual neighbors).” The manuscript lacks a rigorous mathematical or temporal definition for this boundary. It remains entirely unclear whether $\mathcal{S}(i)$ is determined by a fixed temporal window (e.g., $\pm N$ seconds), shared structural annotations (e.g., all clips strictly within the same Step), or a predefined semantic similarity threshold. Without a precise algorithmic definition, reproducibility is severely compromised. How exactly does the framework distinguish between a “contextual neighbor” that warrants adaptive down-weighting and a “true negative” from an adjacent but distinct phase that should be strictly separated?
Q3: Insufficient Visual and Quantitative Evidence for Hierarchy Tree Preservation. The authors emphasize that HyperVLP successfully preserves the intrinsic hierarchy tree of surgical workflows. However, the primary interpretability analysis provided (Figure 2) only demonstrates the radial distance distribution, merely proving that broad Phases are closer to the origin than fine-grained Steps. This 1-dimensional analysis completely fails to validate the actual topological branching and cross-level semantic consistency ，whether a specific Step is geometrically subsumed exclusively within the correct Phase’s entailment cone.
- Please rate the clarity and organization of this paper
Good
- 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.
(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?
This paper presents a hyperbolic video-language pretraining framework (HyperVLP) for modeling hierarchical surgical workflows. The shift from flat Euclidean embeddings to a geometry-aware representation space is conceptually meaningful and aligns well with the inherent structure of surgical procedures. However, several concerns regarding methodological grounding and empirical validation of the central claims prevent a higher score.
Strengths：
Hierarchical Reframing via Hyperbolic Geometry: The paper introduces a principled reformulation of surgical video-language pretraining by embedding representations in a Lorentz hyperbolic space. This design is well-motivated, as hyperbolic geometry naturally supports tree-structured data and enables low-distortion modeling of semantic containment across phases, steps, and actions.
Awareness of Structural False Negatives: The authors identify an important limitation in standard contrastive learning, namely the presence of structurally related samples being incorrectly treated as negatives. The proposed geometry-aware contrastive formulation introduces a more nuanced alignment strategy by down-weighting such samples, which is practically relevant for surgical workflows.
Consistent Empirical Gains: The method demonstrates stable improvements across multiple benchmarks, including cross-procedure and cross-center settings. These results suggest that the framework has practical value and contributes to more robust representation learning in surgical domains.
Weak：
Heuristic Nature of the Contrastive Mechanism: The proposed geometry-aware contrastive loss relies on a similarity-based reweighting term (αᵢₖ), which appears closely related to existing false-negative mitigation strategies in contrastive learning. As such, it functions more as a heuristic adjustment rather than a principled mechanism for enforcing hierarchical structure. This creates a gap between the stated goal of hierarchy preservation and the actual optimization process.
Lack of Theoretical or Structural Guarantees: The framework does not provide a clear justification that the proposed reweighting scheme can reliably recover or preserve a discrete hierarchical topology. In the presence of noisy supervision or spurious similarities, the mechanism may fail to distinguish true negatives from structurally related samples, potentially leading to degraded structure modeling.
Ambiguity in the Definition of Structural Relationships: The formulation of structurally related samples (𝒮(i)) lacks a precise operational definition, particularly regarding temporal boundaries or selection criteria. This ambiguity affects reproducibility and makes it unclear how the model differentiates between contextual neighbors and true negatives in practice.
Insufficient Evidence for Hierarchy Preservation: While the paper claims to preserve hierarchical structure, the provided analysis is limited to radial distance distributions, which only capture coarse semantic ordering. There is no direct evidence demonstrating correct parent–child containment or hierarchical topology, leaving the central claim insufficiently supported.
Final Assessment
The proposed framework is well-motivated and practically effective, with consistent empirical improvements and a coherent design. However, the concerns regarding limited methodological novelty, heuristic formulation, and insufficient validation of hierarchy preservation weaken the overall contribution.
I am open to raising the score if the authors can better clarify the structural assumptions, strengthen the connection between the method and hierarchical modeling, and provide more convincing evidence supporting their core claims.
- 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 for their constructive feedback. We address the major concerns below.
[R1/2/3] Euclidean vs. Hyperbolic Gains & Evaluation Connection First, our comparisons are not limited to flat VLP. HecVL and PeskaVLP are Euclidean hierarchical surgical VLP baselines: HecVL learns hierarchy-specific Euclidean embeddings at clip/phase/video levels, PeskaVLP adds hierarchical knowledge augmentation and a procedure-aware objective. Tab. 1/2 compare HyperVLP against them under identical downstream protocols. Second, Tab. 3 disentangles each component: from the flat VLP baseline (line 1), adding hyperbolic geometric contrastive (line 2) and entailment losses (line 3/4/5) progressively improves all four benchmarks. Finally, Fig. 2 shows that abstract phases lie closer to the origin, while fine-grained steps lie farther away in hyperbolic space, confirming that HyperVLP geometrically encodes the abstract-to-specific ordering of surgical workflows.
[R1/2/3] Hierarchy labels & Parent-child & S(i) definitions The hierarchy labels are derived from the same SVL dataset used by HecVL/PeskaVLP: clip-level narrations come from ASR transcripts, phase/video-level texts come from keystep summaries, and parent-child relations are defined by temporal containment rather than learned semantic similarity. Specifically, a child clip/step is assigned to the parent phase/video segment that temporally contains it which is not based on a fixed temporal window or a text-similarity threshold. Based on this hierarchy metadata, S(i) contains non-identical samples from the same surgical video id and the same hierarchy level/context. These samples are not positives, but they are procedurally related and should not be treated as hard negatives. Samples outside this structural group, such as those from different procedures, remain ordinary negatives. Since the ASR transcripts are inherited from the same SVL dataset, ASR noise is shared by all SVL-based baselines.
[R2/3] Hierarchy Evidence Figure 2 shows that geodesic distance to the origin reflects semantic specificity: phase-level videos/texts are closer to the origin than step-level videos/texts, and this consistent cross-modal radial ordering supports hierarchy-aware organization.
[R1/2] Fairness & Parameter Stability & Uncertainty All baselines follow PeskaVLP’s downstream protocol (same SVL source, splits, zero-shot/linear-probing settings). Hyperparameters were selected once following hyperbolic entailment literature and fixed across all benchmarks. Following prior work, we report single-run results and will release repeated-run statistics if accepted.
[R3] Theoretical Grounding L_GeoHCL does not reconstruct a tree from learned similarity; the hierarchy is given by SVL temporal metadata, and L_GeoHCL only prevents false negatives within it. The primary structural constraints are cone-based entailment losses (L_inter, L_intra), enforcing parent-child containment in hyperbolic space. Two-stage training (Stage I: cross-modal alignment; Stage II: hyperbolic layers fine-tuned with encoders frozen) reduces collapse risk.
[R2] More evaluations MultiBypass140 provides step-level annotations, and we agree that step prediction and parent-child retrieval are valuable for evaluating hierarchical consistency. We have conducted step-level evaluation on MultiBypass140 and observed trends consistent with the phase-level results. We omitted these results due to space constraints and to keep the main evaluation protocol aligned with PeskaVLP for fair comparison. Since rebuttal rules don’t allow new experimental results, we will include these analysis in the released project if accepted.
[R1/2] Notations We thank the reviewers for catching these issues. The second V_c should be T_c, and “child-level features v and t” is correct. W_v and W_t are learnable projections to the tangent space before the Lorentz exponential map, fine-tuned in Stage II with frozen encoders.
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 submission presents a well-motivated methodological contribution on hierarchical surgical video-language pre-training, with clear practical gains across multiple cross-procedure and cross-center benchmarks. The main strengths are the coherent use of hyperbolic geometry to model workflow hierarchy, the combination of geometry-aware contrastive learning with entailment-based constraints, and the consistently improved zero-shot and few-shot phase recognition results relative to prior general-purpose and surgical baselines. The paper is technically interesting and addresses an important problem in surgical representation learning. The decisive concern is that the manuscript’s strongest claim - namely that the proposed hyperbolic formulation preserves hierarchical surgical structure more effectively than Euclidean alternatives - is only indirectly supported by the current evidence. After checking the manuscript, this concern is well supported because the evaluation is centered on phase recognition, the hierarchy analysis is limited to radial-distance distributions, and there is no matched Euclidean counterpart with identical hierarchical supervision to isolate the effect of hyperbolic geometry itself. In addition, some methodological details remain insufficiently specified, especially the operational definition of structurally related samples in the contrastive loss. At the same time, I assign limited weight to the request for formal theoretical guarantees, since this appears to stem from an expectation that is stronger than what is normally required for an empirical MICCAI methodology paper.
- 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 initial reviews were generally positive about the motivation, relevance, and empirical performance, but raised several important concerns. R1 and R2 both questioned whether the observed gains can be attributed specifically to hyperbolic geometry rather than to additional hierarchy-aware supervision. The meta-review similarly identified this as the decisive concern. R2 also noted that the evaluation is mainly based on flat phase recognition, which does not directly test hierarchical containment. R1 raised additional concerns about comparison fairness, dataset/hierarchy construction, parameter stability, uncertainty estimates, and notation. R3 was more positive overall, but also pointed out that the definition of structurally related samples and the evidence for topology preservation were insufficiently specified. The rebuttal addressed a meaningful part of these concerns. It clarified that HecVL and PeskaVLP are Euclidean hierarchical surgical VLP baselines evaluated under the same downstream protocol, that Table 3 compares the proposed components with a flat VLP baseline, and that structurally related samples S(i) are defined using SVL temporal metadata rather than learned semantic similarity or fixed text thresholds. It also clarified that parent-child relations are based on temporal containment, that ASR noise is shared with SVL-based baselines, and that the main structural constraints are the inter-modal and intra-modal entailment losses rather than the contrastive reweighting alone. These clarifications reduce concerns about baseline fairness, reproducibility of S(i), and the interpretation of the contrastive objective. However, some limitations remain. The submitted paper still lacks a fully matched Euclidean counterpart using the same hierarchical losses, so the specific benefit of negative curvature is not completely isolated. The hierarchy analysis remains mostly based on radial distance distributions, which support abstract-to-specific ordering but do not directly demonstrate correct parent-child retrieval, cone containment, or tree topology preservation. The paper also reports single-run results without uncertainty estimates, and the parameter selection/stability discussion remains limited. The rebuttal mentions step-level evaluation trends on MultiBypass140, but these results are not part of the submitted manuscript and are therefore not used as decisive evidence. The post-rebuttal opinions were mixed but overall support acceptance. R2 changed to Accept, judging that the rebuttal clarified the main concerns sufficiently and that the work has technical merit and relevance to surgical AI. R1 remained Reject, but their post-rebuttal text acknowledges that the authors clarified the Euclidean baseline fairness and the S(i) definition; the remaining objection focuses mainly on missing additional hierarchy-specific validation. R3 did not provide an updated post-rebuttal opinion in the provided material, so the initial Weak Accept is retained as available evidence. In weighing these positions, I assign substantial weight to the remaining methodological limitation, but do not consider it fatal given the coherent formulation, competitive empirical results, and rebuttal clarifications.
Meta-review #2
- 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 paper proposes a hyperbolic surgical video-language pre-training framework that models workflow hierarchy through geometry-aware contrastive learning and entailment-based constraints. The conceptual interest of the hyperbolic formulation and the consistent improvements across zero-shot and few-shot phase recognition benchmarks were acknowledged. The decisive concern, raised by two reviewers, is that the paper’s strongest claim, namely that hyperbolic geometry preserves hierarchical surgical structure more effectively than Euclidean alternatives, is only indirectly supported. The evaluation is centered on flat phase classification; the hierarchy analysis is limited to radial-distance distributions; and there is no matched Euclidean counterpart with identical hierarchical supervision to isolate the effect of hyperbolic geometry itself. The rebuttal argues that two cited Euclidean hierarchical baselines play this role, but they do not constitute a matched ablation in which only the geometry differs. One reviewer maintained Weak Reject, explicitly noting that the additional empirical validations needed to substantiate the hierarchical claim are not yet in the manuscript; the other Accept vote also flagged residual uncertainty about whether the gains arise specifically from hyperbolic geometry. The work is technically interesting, and a future iteration with the missing controls could be strong, but in its current form, the central claim is not adequately supported.
Meta-review #3
- 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.
Good work. The authors should add more discussion in the final version to address R1’s concerns.