{"id":"0c952604-a5a7-4900-b3b9-1ceba50703e8","arxiv_id":"2506.01980","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"C2E, a masked-autoencoder-style surgical foundation model with a compression encoder and entropy-maximizing decoder, reports gains over several baselines but with a theory-implementation gap and no released code.","lead":"This paper introduces C2E, a self-supervised surgical vision model pretrained on 0.78 million unlabeled frames, and reports improved fine-tuned performance on phase recognition, tool-tissue interaction, segmentation, and polyp diagnosis. A generalist would read it to see whether compression-based self-supervision can reduce the need for expensive expert annotations in surgical AI.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Paper's own Section II-D states training runs as a standard MAE pipeline, so the entropy-maximization and compression objectives in Eq. 3 and Eq. 10 are never actually optimized; the central C2E mechanism is disconnected from the implemented loss.","rationale":"I agree with the reader's weakest assumption. The paper's Section II-D is explicit that the loss is standard MAE reconstruction, so the entropy-maximizing and Langevin machinery in Section V is never connected to an optimized objective. The theorems also contain directional errors (Theorem 1 states an equality that is only an asymptotic limit, and Theorem 3 mislabels H(Z) as a higher bound when the cited inequality makes it the lower side). The private dataset and absence of code make the empirical claims impossible to verify independently, and the comparison to ViT in Table VI is not a controlled ablation of the proposed decoder. A conditional request for code might be appropriate if the central mechanism were merely unverified, but here the internal contradiction between the described training and the claimed objective means the central claim as written is unsupported. Therefore REJECT is the appropriate verdict.","tokens_in":19015,"tokens_out":3329,"duration_ms":33711,"concrete_test":"Obtain the official code and configuration (or, if unavailable, a faithful reimplementation) and inspect the training graph. Specifically: (1) Does the decoder update in Eq. 10 appear in the forward and backward pass, or is it only used for sampling? (2) Is the loss exactly MAE/L1 on masked patches with no entropy or log-determinant term? If the loss is plain MAE, run an ablation with a standard MAE decoder of identical encoder and data; if downstream metrics (Cholec80 phase accuracy, CholecT45 mAP, CholecSeg8k IoU, PolypDiag accuracy) remain within one standard deviation, the claimed entropy-maximization mechanism is not load-bearing.","verdict_should_be":"REJECT","load_bearing_attack":"The most load-bearing concern is that the paper's own training description contradicts its theoretical contribution. Section II-D states: 'The loss function minimizes the mean average error of the predicted images and the original images, making the overall training run as a standard MAE [17] pipeline.' The methods sections II-B and II-C claim a compression encoder that maximizes Kolmogorov-complexity difference (Eq. 3) and an exploration decoder that maximizes entropy via Langevin sampling (Eq. 10). But a plain MAE reconstruction loss contains neither the log-determinant term -1/2 ln|Σ_Z| from Eq. 3 nor any gradient signal that trains the energy function E(z) or temperature kT(h) in Eq. 10; those terms appear in no loss term and no weight update. Thus, if the implemented training is standard MAE, the trained encoder is not the one justified by the theorems, and C2E reduces to an architectural variant of MAE. The theoretical section does not repair this: Theorem 1's proof cites Cover and Thomas Theorem 14.3.1 and gives E[1/n K(X)] → H(X), but the stated equality H(X)=K(X) is not established; Theorem 3 calls H(Z) a 'higher bound' of K(X) when Eq. 13 places H on the lower side. With no code, no training details, and 531k private MGH images, the reported gains cannot be attributed to the claimed mechanism rather than to data scale or standard MAE pretraining. This disconnect is the single point on which the central claim hinges.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Compress-to-Explore (C2E), a self-supervised pretraining framework for surgical video understanding. The method claims a compression encoder that maximizes a Kolmogorov-complexity difference objective and an exploration decoder based on entropy-maximizing Langevin sampling, pretrained on 0.78M unlabeled surgical images from 2122 procedures. The pretrained encoder is then fine-tuned on phase recognition, action-triplet classification, segmentation, polyp diagnosis, and few-shot phase recognition. The central claims are that C2E produces compact, disentangled representations and improves downstream accuracy relative to ViT/MAE-style baselines, thereby serving as a surgical visual foundation model.","tokens_in":19340,"tokens_out":6495,"duration_ms":66830,"significance":"If the empirical results were reproducible and the mechanism were actually implemented, the scale of pretraining (0.78M frames, 2122 surgeries) and the breadth of downstream tasks would be a useful contribution to surgical computer vision. The paper also includes few-shot evaluations and representation visualizations, which are appropriate for a foundation-model claim. However, the theoretical core supporting the C2E mechanism is not sound, the implemented loss appears to be standard MAE reconstruction, and the empirical comparisons do not consistently show the claimed improvements. The paper does not release code or the private MGH dataset, so the reported gains cannot currently be attributed to the proposed mechanism rather than to data scale or an architectural variant of MAE. The significance as stated, namely a new compression- and entropy-based self-supervised principle, is therefore not established.","major_comments":[{"comment":"The paper explicitly states that 'the loss function minimizes the mean average error of the predicted images and the original images, making the overall training run as a standard MAE [17] pipeline.' None of the terms in Eq. 3, such as -1/2 ln|Σ_Z| and the dimension-change term, appears in this loss, and Eq. 10's energy, temperature kT(h), and Langevin noise are likewise absent from any weight update. Consequently, the trained encoder is not the one whose compression and entropy-maximization properties are proved in Section V. This is a load-bearing inconsistency: the central claim that C2E 'leverages Kolmogorov complexity' and 'uses entropy-maximizing decoders' is disconnected from the implemented training objective.","section":"Section II-D"},{"comment":"Theorem 1 states that the entropy H(X) is the Kolmogorov complexity K(X), but the proof establishes only the asymptotic relation H(X) ≤ E[(1/n)K(X)] ≤ H(X) + ... for finite alphabets, i.e., that the expected normalized Kolmogorov complexity converges to H(X). This does not imply equality of H(X) and K(X) for individual embeddings. In addition, Theorem 3 refers to H(Z) as a 'higher bound' of K(X), while Eq. 13 places H(X) as a lower bound of the expected complexity; the inequality is used in the wrong direction. These errors undermine the theoretical foundation of the compression encoder.","section":"Section V, Theorems 1 and 3"},{"comment":"The algebraic derivation of Eq. 4 is not valid as written. From Z = SVD, the simplification Z(Z^T Z)^{-1} = S V^{-1} D requires orthogonality and invertibility assumptions that are not stated and are not satisfied by the rectangular matrix Z ∈ R^{N,C}; the computation also drops D^{-1} and misplaces the transpose structure. Separately, Theorems 2 and 5 invoke the concavity of ln|A| on positive definite Hermitian matrices and assert that 'Z is a positive definite Hermitian matrix', but the paper's own notation defines Z as an N×C matrix of hidden states. The matrix-analysis theorems therefore do not apply as stated, leaving Eqs. 3-5 without a valid justification.","section":"Section II-B, Eq. (4), and Section V, Theorems 2 and 5"},{"comment":"The maximum-entropy solution is misidentified in Eqs. 6-8. Under a constraint on E[E(z)], the Lagrange multiplier λ is the inverse temperature, and the solution is p(z) ∝ exp(-λ E(z)); it is not p(z) ∝ exp(-E(z)/E[E(z)]). Writing E[E(z)] = kT(h) in Eq. 8 conflates the expectation of the energy with the temperature parameter, yet this identification is used directly in the Langevin sampling of Eq. 10. The exploration decoder's theoretical grounding is therefore also unsupported.","section":"Section II-C, Eqs. (6)-(8)"},{"comment":"The reported results do not consistently support the claim of improved performance. In Table II, C2E (92.5±6.9) is statistically indistinguishable from SurgFormer (92.4±6.4), and its precision is lower than several baselines. In Table III, C2E's mAPit (42.3±0.9) is below MT4MTL's (43.1±2.0), even though the text claims better average precision across the action-triplet metrics. More importantly, no ablation controls for the 531,565 private MGH images or for the specific architecture changes against a standard MAE trained on the same data; the only ViT comparison in Table VI uses a previously published model rather than a same-data, same-compute MAE baseline. Without such controls, the reported gains cannot be attributed to the C2E mechanism.","section":"Section III, Tables II-VI"}],"minor_comments":[{"comment":"There are numerous typos, including 'Massachusset' in the author affiliation, 'adoped' in Section III-D, 'few-short learning' in the contributions list and Section III-F, and 'Precession' instead of 'Precision' in the Table II header.","section":"Throughout"},{"comment":"The entry '58.8 ±11.' is incomplete; the standard deviation is missing its decimal digits.","section":"Table VII"},{"comment":"The naming is inconsistent: 'VIT' is used in Figure 3 and Table VI while 'ViT' is used elsewhere; please unify.","section":"Section III-F and Figure 3"},{"comment":"The symbol E is used both for the energy function and for the expectation operator, e.g., E[E(z)] in Eqs. 6-8 and Eq. 10; this is confusing and should be disambiguated with different symbols.","section":"Section II, notation"},{"comment":"The claim that Z0 is a 'disentangled and sparse normal distribution' is not supported by any quantitative evaluation in the paper; the t-SNE visualizations in Section III-F are not a substitute for a measured sparsity or disentanglement metric.","section":"Section II-A"}],"recommendation":"reject","confidential_remarks":"The paper's central contribution depends on a private 531,565-image dataset and no code release, while the theoretical claims are internally inconsistent and the implemented objective is described as standard MAE. I do not see a path to acceptance within the current scope. A substantially rewritten paper that drops the unsupported theory and positions the contribution as an empirical study of a MAE-style architecture pretrained at surgical scale could be a different, potentially publishable manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has one real strength and one load-bearing flaw. The strength is empirical: the authors assembled a 0.78M-image pretraining corpus from 2122 surgeries and evaluated the resulting encoder on four downstream tasks against external benchmarks. The few-shot results in particular look substantially better than the VIT baseline, and the action-triplet mAPivt gain (41.8 vs 37.1) is not trivial. If the numbers are reproducible, this is a useful data point for surgical SSL.\n\nThe flaw is that the paper's own methods section contradicts its theoretical contribution. Section II-D states the loss \"minimizes the mean average error of the predicted images and the original images, making the overall training run as a standard MAE pipeline.\" That means the log-determinant compression objective in Eq. 3 and the entropy-maximizing Langevin decoder in Eq. 10 are never actually optimized. The theorems in Section V justify a training procedure that the paper does not run. This is not a minor gap—it severs the connection between the theory and the reported results. The theorems themselves are also sloppy: Theorem 1 claims H(X)=K(X) but the proof only shows asymptotic closeness, and Theorem 3 calls H(Z) a higher bound of K(X) when the paper's own inequality places it on the lower side. These are not just typos; they suggest the theoretical machinery was written after the fact rather than derived from the implementation.\n\nThere are other soft spots. No code, no training details, and a large private dataset (531k MGH images) mean the empirical results cannot be checked. Some reported gains are within noise, such as phase accuracy 92.5 vs 92.4. The representation analyses (t-SNE, saliency maps) are suggestive but not quantitative.\n\nWho gets value from this paper? Researchers working on surgical foundation models might benefit from the dataset composition and evaluation protocol, even if they ignore the theory. But as a claim about compression and entropy maximization improving representation learning, it does not hold together in its current form.\n\nIf I were the editor, I would send it to peer review anyway, because the empirical scale is real and the contradictions deserve to be exposed in review. But I would expect major revision or rejection. The authors need to either show how the MAE loss implements the proposed compression and entropy objectives, or drop the theory and present this as a careful study of MAE pretraining at scale.","headline":"Potentially useful large-scale surgical pretraining study undermined by a disconnect between its theoretical framing and its implemented MAE-style loss.","tokens_in":19870,"tokens_out":2425,"would_cite":false,"duration_ms":25174,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A self-supervised compression objective trained on 0.78 million unlabeled surgical frames produces an encoder that fine-tunes to improved results across surgical phase recognition, action triplet detection, segmentation, and polyp…","keywords":["surgical foundation model","self-supervised learning","Kolmogorov complexity","entropy maximization","masked autoencoder","surgical phase recognition","surgical action triplets","few-shot learning"],"falsifier":"Train the same architecture with the same data, masking, and training schedule but replace the decoder and losses with a plain MAE decoder and L2 or MAE reconstruction; if phase, action-triplet, segmentation, and polyp metrics are statistically indistinguishable from C2E's, the compression and entropy machinery is not carrying the reported gains. Conversely, the paper's claim would be supported by showing that removing the entropy-maximizing Langevin decoder or the dimension-reduction projections measurably degrades downstream accuracy.","tokens_in":18773,"feed_emoji":"🩺","tokens_out":5196,"duration_ms":47175,"temperature":0.7,"pith_summary":"The paper proposes Compress-to-Explore (C2E), a self-supervised pretraining method that trains a vision transformer encoder on 0.78 million unlabeled images drawn from 2122 minimally invasive surgeries. The idea is to make the encoder learn representations by compressing images through stages that reduce Kolmogorov complexity, while a decoder built on entropy-maximizing energy functions reconstructs clinically relevant detail. The paper claims that the resulting encoder, fine-tuned on downstream tasks, outperforms or matches existing surgical vision baselines in phase recognition, instrument-tissue action triplets, segmentation, and polyp diagnosis, and that it generalizes with very few labeled videos on new procedure types. A sympathetic reader would take the central claim to be that an unlabeled-data compression objective, rather than task-specific labels, can carry a general surgical visual foundation model.","feed_headline":"Surgery video model trained without labels beats four task benchmarks","feed_subtitle":"Pretrained on 0.78M unlabeled surgical frames, the C2E encoder lifts phase, action, segmentation, and polyp diagnosis scores.","key_machinery":"The load-bearing object is the compression encoder: each layer removes a constant dimension and applies a shrinkage step derived from the objective $\\max_\\theta\\,-H(Z)+H(I)$, which the paper connects to Kolmogorov complexity through the asymptotic identity $E[\\frac{1}{n}K(Z^n)]\\to H(Z)$. The gradient update uses an SVD-based form $Z_{i+\\frac12}=Z_{i+1}-\\beta(S_{i+1}V_{i+1}^{-1}D_{i+1})$ with a residual-like projection bypass, so each layer is meant to solve a concave entropy-maximization subproblem. The decoder inverts this with a temperature-dependent Langevin sampler, where the temperature is estimated from a conditional slice of the hidden state. In the actual pretraining, however, the stated loss is mean absolute error between reconstructed and original patches, run as a standard masked autoencoder pipeline.","core_discovery":"On the paper's own terms, C2E establishes that maximizing the entropy difference between input and a dimension-reduced hidden state, equivalently minimizing Kolmogorov complexity under the identification of complexity with entropy, yields an encoder whose latent representations are compact and disentangled by surgery type, phase, and instrument. The reconstruction side uses a decoder derived from the closed-form entropy-maximizing distribution $p(z)\\propto e^{-E(z)/E[E(z)]}$ sampled via Langevin dynamics, intended to recover perceptually fine details such as tool textures and organ boundaries. The paper reports, with the encoder used as a backbone, 92.5% accuracy on Cholec80 phase recognition, 41.8 mAP on action triplet recognition, 0.74 IoU on CholecSeg8k segmentation, and 94.0% accuracy on polyp diagnosis, and substantially higher few-shot phase accuracy than a vision transformer pretrained on endoscopic images.","pith_inferences":["The reported comparisons mostly keep the downstream decoder fixed while swapping the backbone, so the incremental gains may come from the larger and more diverse pretraining corpus rather than from the compression theory; a controlled experiment pretraining plain MAE on the identical corpus would separate the two.","If the effective objective is standard MAE, C2E can be read as evidence that scaling up diverse surgical videos with a masked-autoencoder objective is sufficient, with the Kolmogorov-complexity framing serving as an interpretation rather than a mechanism.","A testable extension is to evaluate C2E on other fine-grained medical video domains, such as endoscopy or retinal surgery, where local texture is decisive; the paper's own logic predicts the largest gains where global and local features must be balanced.","The temperature-conditional decoder suggests a route to per-image uncertainty estimates in downstream predictions, though the paper does not attempt this."],"forward_implications":["If C2E's compression objective is what drives its representations, unlabeled video libraries become the main resource for building surgical foundation models rather than labeled benchmarks.","Fine-tuning C2E on as few as two or eight videos yields phase recognition gains of roughly 8 to 26 points over a ViT baseline on colorectal procedure types, suggesting new procedures can be modeled with minimal annotation.","Because one pretrained encoder helps classification, triplet detection, segmentation, and diagnosis, downstream systems can share a single backbone instead of task-specific pretraining.","The claimed disentanglement of surgery type, phase, and instrument in the latent space could enable interpretable monitoring, with attention heads tracking local details and their aggregation composing global context."],"supporting_citations":[{"why":"Supplies the masked autoencoder pipeline and masking rule that C2E inherits, plus the reconstruction-style pretraining loop.","marker":"[17]"},{"why":"Provides the entropy-to-Kolmogorov complexity identity and the multivariate normal entropy formula that ground the compression objective.","marker":"[64]"},{"why":"Supplies the Cholec80 dataset and phase labels used for the surgical workflow classification benchmark.","marker":"[70]"},{"why":"Provides the ViT baseline pretrained on endoscopic images that C2E is compared against in ablations and few-shot tests.","marker":"[19]"},{"why":"Defines the CholecT45 benchmark and metrics for surgical action triplet recognition used in the action classification experiments.","marker":"[86]"},{"why":"Supplies the CholecSeg8k segmentation dataset used to evaluate IoU against prior segmentation methods.","marker":"[52]"},{"why":"Supplies the PolypDiag dataset and protocol for the intraoperative polyp diagnosis task.","marker":"[106]"},{"why":"Provides the Rendezvous attention method adopted as the downstream decoder for action triplet recognition.","marker":"[88]"}],"fun_headline_variants":["No-label surgical AI (C2E) beats supervised on four tasks","Self-supervised C2E wins on 4 surgical benchmarks, no labels needed","C2E: label-free surgical AI wins on four benchmarks","Entropy-max compression yields self-supervised surgical model","C2E's entropy trick: no labels, better surgical vision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pretraining that produced the reported results is the same network described by the compression and entropy-maximization objectives, rather than a standard masked-autoencoder reconstruction loss with the theoretical machinery playing no functional role.","fun_headline_variants_meta":{"raw":{"variants":["No-label surgical AI (C2E) beats supervised on four tasks","Self-supervised C2E wins on 4 surgical benchmarks, no labels needed","C2E: label-free surgical AI wins on four benchmarks","Entropy-max compression yields self-supervised surgical model","C2E's entropy trick: no labels, better surgical vision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001191,"raw_usage":{"total_tokens":4922,"prompt_tokens":958,"completion_tokens":3964,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":3873}},"tokens_in":574,"tokens_out":3964,"duration_ms":27314,"temperature":1.0,"reasoning_tokens":3873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:54:41.144033+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture with the same data, masking, and training schedule but replace the decoder and losses with a plain MAE decoder and L2 or MAE reconstruction; if phase, action-triplet, segmentation, and polyp metrics are statistically indistinguishable from C2E's, the compression and entropy machinery is not carrying the reported gains. Conversely, the paper's claim would be supported by showing that removing the entropy-maximizing Langevin decoder or the dimension-reduction projections measurably degrades downstream accuracy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the entropy-to-Kolmogorov complexity identity and the multivariate normal entropy formula that ground the compression objective."},{"cited_title":"EndoNet: A deep architecture for recognition tasks on laparoscopic videos,","cited_arxiv_id":null,"evidence_quote":"Supplies the Cholec80 dataset and phase labels used for the surgical workflow classification benchmark."},{"cited_title":"CholecTriplet2021: A benchmark challenge for surgical action triplet recognition,","cited_arxiv_id":null,"evidence_quote":"Defines the CholecT45 benchmark and metrics for surgical action triplet recognition used in the action classification experiments."},{"cited_title":"Contrastive transformer-based multiple instance learning for weakly supervised polyp frame detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the PolypDiag dataset and protocol for the intraoperative polyp diagnosis task."},{"cited_title":"Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos,","cited_arxiv_id":null,"evidence_quote":"Provides the Rendezvous attention method adopted as the downstream decoder for action triplet recognition."}],"review_version":1}