acceptodds
Withdrawn from ICLR 2027

SurgicalMamba: Dual-Path SSD with State Regramming for Online Surgical Phase Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Online surgical phase recognition labels each frame of an operation from past frames alone. Because procedures run for hours, the cost of each prediction can- not grow with elapsed time. Structured state-space duality (SSD) meets that constraint, but is not well matched to surgical video in two ways. The same views recur through an operation, so a scene is written over its earlier occur- rence and afterwards only age separates the two. Phase durations differ by an order of magnitude, so a decay fast enough to clear a phase that passes in sec- onds erodes one that runs for minutes. The rate does vary from frame to frame, but on contexts this long it is not reliably learned into a forgetting schedule. State regramming applies an input-conditioned orthogonal rotation to the state handed across chunk boundaries, so where a frame is written also depends on what has passed since, and two occurrences of the same view are held apart when different phases intervene. Intensity-modulated stepping supervises a time- warp of the discretization step toward the annotated phase boundaries, clear- ing the state where a phase ends so that the decay within phases can lean to- ward retention. Both preserve SSD’s N -semiseparable structure and O(d) per- frame cost. On seven public benchmarks SurgicalMamba matches or exceeds the state of the art in online accuracy and phase-level Jaccard (94.6%/82.7% on Cholec80, 89.5%/68.9% on AutoLaparo) at 312.88 fps on a single GPU. Code :https://anonymous.4open.science/r/Surgical-Mamba-8930

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
surgical phase recognition, surgical video analysis, state space models, Mamba, structured state-space duality (SSD)

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RKZrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SurgicalMamba for online surgical phase recognition. Built on Mamba2/SSD, it introduces three main changes. First, it applies content-conditioned orthogonal transformations to recurrent states at chunk boundaries, a process called state regramming. Second, it uses boundary supervision to adjust the discretization step size and control forgetting. Third, it uses separate paths for long- and short-term modeling, each with a different state-reset policy.

Critical strengths
  1. The model design addresses the problems faced: the appearance of similar images at different stages and the stages needed to be forgotten correspond to the model design in the article.
  2. Content-conditioned state rotation is the most promising methodological contribution. Allowing intervening content to modify the orientation of stored information is an interesting direction for recurrent memory.
Critical weaknesses
  1. The paper does not clearly establish its methodological advance over the backbone. Mamba2 already uses an input-dependent step size to control forgetting. The proposed intensity modulation mainly adds boundary supervision and a constrained reparameterization of that mechanism. Likewise, the dual-path design combines long- and short-term processing through different reset policies and cross-path conditioning. Besides, I feel this paper leans more toward an engineering innovation rather than a methodological one. These design choices may be useful, but the paper does not show why they are better than simpler alternatives. One important comparison would be to apply the same boundary supervision directly to Mamba2’s original step-size mechanism. Strong results on surgical phase recognition alone do not establish a methodological advance over selective state-space models.

  2. The experiments do not isolate what makes state regramming effective, even though it is the paper’s main novelty claim. Removing the rotation module shows that it helps the full system. However, this does not show whether the gains depend on content-conditioned rotation subspaces and history-dependent transformations. Fixed rotations or generic learned state mixing might provide similar benefits. The norm-preservation and conditional kernel analyses establish structural properties, but they do not directly show reduced memory interference. Targeted comparisons or direct evidence of this mechanism are still needed. Without them, it is unclear whether the gains come from the proposed principle or simply from adding another architectural component. This weakens the case for accepting the paper as a general methodological contribution.

Questions for authors

The following questions concern evidence and clarification that could materially change my assessment. I am open to rerating if they can be explained properly in rebuttal.

  1. Intensity modulation: Under identical boundary supervision, how does the proposed modulation compare with directly supervising Mamba2’s original selective step size or using a simple boundary-conditioned forgetting gate? If such results are unavailable, can the authors clarify what advantage of the proposed parameterization is established independently of the additional supervision?

  2. State regramming: How do the proposed content- and history-conditioned rotations compare with simpler alternatives, such as fixed or input-independent rotations, content-dependent rotations in fixed planes, and nonorthogonal state mixing with the same parameter count? Which results show that the gains come from the specific rotation design, rather than simply adding a state transformation?

Reviewer_YWuDOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces SurgicalMamba, a dual-path SSD architecture for online surgical phase recognition. Its main contributions are content-conditioned state rotations to distinguish recurring visual patterns and intensity-modulated stepping to encourage forgetting near phase boundaries. Experiments across seven surgical datasets report competitive performance, supported by theoretical analysis, ablations, and streaming efficiency measurements.

Critical strengths
  1. The paper addresses a real, well-defined problem with a coherent streaming-compatible design. The two failure modes of SSD are clearly motivated, and both proposed mechanisms directly target them while preserving the chunked SSD structure. The structural result that the rotation can be absorbed into the B/C projections, preserving N-semiseparability and fixed per-frame cost, gives the method a principled foundation rather than making it an ad hoc architectural stack. This is the main reason the core idea remains worth pursuing despite the empirical issues.
  2. The reported evaluation is broad and mechanistically specific. The paper covers seven benchmarks, includes ablations, and provides an error analysis that distinguishes the effects of Z and λ: removing Z concentrates errors on already-seen phases, while removing λ concentrates errors on neighboring closing phases. If the reported protocol is accurate, this would be a meaningful contribution.
Critical weaknesses
  1. Significant discrepancies between manuscript and released code: The Cholec80 and M2CAI16 training sets overlap their validation sets in the released implementation, contradicting the reported disjoint splits. The released training configuration also differs in clip length and loss weights. Furthermore, the intensity-network input is not detached as described in the paper, weakening the claimed isolation of boundary supervision. These inconsistencies reduce confidence in the experimental conclusions.

  2. Incomplete reproducibility and comparative validation: The released repository provides training support for only three of the seven reported benchmarks and lacks scripts for reproducing the ablations and statistical tests. Moreover, no controlled comparison with Mamba-3 is provided to establish the incremental benefit of the proposed rotation mechanism. There are also latent implementation issues: SurgicalMamba.step ignores the use_fast_path/use_slow_path flags; OnlineSession shares one chunk counter between fast and slow paths, which is incorrect if their chunk sizes differ; and _reset_fast_states does not zero y_sum_fast when clip_len is not a multiple of the fast chunk size.

Questions for authors
  1. Can the authors explain the discrepancies between the manuscript and released code regarding dataset splits, clip length, loss weights, and gradient detachment? Which exact configurations produced the reported results?
  2. Can the authors provide experimental results using independent validation sets, along with the configurations and evaluation scripts required to reproduce the main tables and ablations?
  3. Can the authors clarify the incremental contribution of state regramming relative to Mamba-3, preferably through a matched experimental comparison?
  4. Why is the intensity-net input not detached, contrary to §3.3, the rotation initialization non-identity, contrary to §3.5?
  5. Once the implementation is reconciled, provide parameter-matched and control baselines: a wider vanilla Mamba2, a static/random chunk rotation, and a Mamba-3-style rotary B/C baseline. These remain necessary to establish that the gains come from content-conditioned rotation rather than added capacity or generic positional encoding.
Reviewer_wDXpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The linked code repository lists the corresponding arxiv-paper with its author in the readme.md file.

I recommend to desk-reject because of anonymity violation.

Critical strengths

NA

Critical weaknesses

NA

Questions for authors

NA

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.