acceptodds
Under review as a conference paper at ICLR 2027

CTKAN: MAKING KOLMOGOROV–ARNOLD NETWORKS THINK OVER SPLINE ACTIVITIES FOR MEDICAL IMAGE SEGMENTATION

Abstract

Neural computation can benefit from evolving internal dynamics rather than relying solely on a single static mapping. Although Kolmogorov–Arnold Networks (KANs) have shown promise in medical image segmentation through learnable spline functions, most KAN-based models evaluate spline responses once and aggregate them immediately, preventing basis-wise contributions from being reconsidered before the layer output is formed. We argue that KANs should think over their spline activities through iterative collaborative thinking, repeatedly refining and coordinating basis contributions before aggregation to strengthen nonlinear representation. We therefore propose CTKAN, a Continuous-Thought Kolmogorov–Arnold Network that introduces neural-dynamics-inspired recurrent evolution into the spline activity space. CTKAN maintains a sample-conditioned state for each output–input–basis triplet and updates these states over multiple thought ticks. Normalized pairwise co-activities provide relational feedback, while terminal states generate residual, sample-adaptive gates over shared spline coefficients, allowing each sample to induce a distinct mixture of spline bases. We integrate CTKAN at two deployment scales into a U-shaped segmentation backbone. Experiments on ultrasound, endoscopic, histopathological, and cardiac MRI benchmarks show that both variants outperform U-KAN, with CTKANMax improving the four-dataset macro-average by 1.29 IoU and 1.03 Dice points. These results demonstrate that iterative collaborative thinking over spline activities improves segmentation performance through sample-adaptive spline modulation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Kolmogorov–Arnold Networks, Continuous Thinking, Iterative Reasoning, Medical Image Segmentation, Dynamic Spline Modulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ojMnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CTKAN, which introduces recurrent updates to spline activities before aggregation in Kolmogorov-Arnold Networks. The method is evaluated on four medical image segmentation datasets and shows improvements over U-KAN and several other variants.

Critical strengths

The idea of updating spline activities before aggregation is interesting. The paper is clearly presented and includes several useful ablation studies. I also appreciate the comparisons with one-shot modulation and post-KAN recurrence, which help explain the proposed design.

Critical weaknesses
  1. My main concern is the dataset splitting protocol. The paper states that the datasets are split at the image level. For ACDC, images from the same patient may appear in different subsets, while CVC-ClinicDB contains correlated frames from the same video sequences. The paper does not clearly explain whether patient-level or sequence-level separation was enforced. This could substantially affect the reliability of the reported results.

  2. Although the authors provide several useful ablations and the implementation of the proposed mechanism, it remains unclear whether the additional gains come specifically from recurrent state evolution or from increased adaptive computation. A stronger non-recurrent baseline with comparable parameters or inference cost would make the contribution more convincing. The synthetic experiment also shows that a simple pre-aggregation MLP can solve the constructed task, which raises questions about the necessity of the recurrent mechanism.

Questions for authors
  1. Were the ACDC and CVC-ClinicDB datasets split at the patient and video sequence levels, respectively? If not, can the authors provide results with group-disjoint splits?

  2. Can the authors compare CTKAN with a parameter-matched or latency-matched non-recurrent adaptive KAN?

  3. Could the authors provide the dataset split assignments and training scripts to support reproducibility?

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

CTKAN adds recurrent spline-coefficient modulation to the first nonlinear channel-mapping layer of selected U-KAN blocks. Image-level statistics drive internal updates before spline aggregation. The best CTKAN variant achieves improvements of 1.29 IoU and 1.03 Dice points over U-KAN across four medical segmentation benchmarks.

Critical strengths
  1. The proposed method uses interactions between spline states to adjust individual basis contributions before they are summed. This allows image conditioning of the edge functions, rather than merely refining the resulting features.

  2. The Post-KAN recurrence controls the main claim of the paper. It shows the impact of the order of operations by showing that recurrence only helps when it acts on the spline coefficients before they are summed.

Critical weaknesses
  1. There is no conclusion on whether 'continuous thought' is actually necessary for this task. Figure 10 and Table 6 report validation accuracy across thought depth and memory time scale (T and L respectively), and what T does depends on L. At L=15 a single tick (81.34/89.59) beats every deeper setting and lands within 0.15 IoU / 0.04 Dice of the selected T=5, L=5 model. The single tick row in Table 2 loses 1.19 IoU, but the paper does not say which L that row uses. From my interpretation, the impact of T would be addressed whether additional thought steps still help once L has been tuned for T=1.

  2. The paper does not explain whether IoU and Dice are calculated for each image and then averaged, or calculated after combining the pixel counts across the dataset. There should be clarity on whether both metrics use the same or different approaches. The results support adapting spline coefficients before they are summed, but it is still unclear whether repeated thought steps are needed.

The main question is whether additional steps still help once L has been tuned for T=1. The paper also needs to clarify why some reported IoU and Dice values are inconsistent when calculated and averaged in the same way.

Questions for authors

See Weaknesses

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

The paper proposes CTKAN, a drop-in replacement for the first spline-linear projection inside U-KAN's tokenized KAN blocks. For each sample, CTKAN pools token statistics of every B-spline basis response, multiplies them by the shared spline coefficients to form a normalized "drive" tensor (from Eq. 3), and then runs T weight-tied recurrent "thought ticks" over a state of the same shape (Eqs. 4–7). At each tick, products of a fixed random set of R_a = 1024 state-coordinate pairs are normalized and passed to an MLP whose output is broadcast additively along the O, I and K axes (Eq. 6). The terminal state produces a residual, sample-global gate that rescales each spline coefficient before the usual aggregation (Eqs. 8–9). Two deployment scales (CTKAN-Light: 2 token stages; CTKAN-Max: all 4) are evaluated on BUSI, CVC-ClinicDB, GlaS and ACDC (2D) with five seeds. CTKAN-Max improves the four-dataset macro-average over U-KAN by 1.29 IoU / 1.03 Dice (Table 1).

Critical strengths
  1. Clear, controlled ablation design. All mechanism variants share the same backbone and the same replacement position (Sec. 4.3), and controls get the same optimizer treatment as the CTKAN dynamics. That makes the variant comparisons in Table 2 easier to interpret than in many architecture papers.

  2. Multi-seed evaluation (five seeds, mean ± SD) with a sealed test set and validation-only hyperparameter selection (Sec. 4.1, Fig. 10).

  3. Near-identity initialization (α₀ = 0.05, small final-layer init) is a sensible design that keeps the module a safe drop-in, and the diagnostics (Fig. 9) show the gates actually move away from identity.

  4. Some evidence of generality beyond medical images (Freiburg Forest, Appendix C).

Critical weaknesses
  1. The core mechanism is close to a sample-conditioned dynamic-weight / squeeze-and-excitation gate.
  • CTKAN computes a per-sample coefficient rescaling from globally pooled basis statistics. Conceptually this is a hypernetwork / dynamic-weight / SE-style gate whose generator is a weight-tied iterative network. The related work covers KAN variants and neural-dynamics models, but not SENet, dynamic convolution / CondConv, HyperNetworks, or weight-tied iterative and implicit models (e.g. DEQ). Without that comparison it is hard to judge what is new beyond "a deeper, recurrent gate generator applied to spline coefficients".
  1. The theoretical motivation is weak (from Appendix E)
  • The proposition (a function of 1ᵀc cannot distinguish c from c + u with u ∈ ker 1ᵀ) is trivially true, but it does not establish a limitation of KAN-based networks. A real layer has O outputs and many tokens, and the next layers also see the residual shortcut. The layer input X, which fully determines c, stays available to any post-KAN or parallel module (an SE block on X, for example). "Post-aggregation modules cannot reconsider basis contributions" is a property of the specific controls built here, not a structural limitation.
  1. Data splitting protocol is leaking information
  • All datasets are split at the image level into 70/15/15 (Sec. 4.1). ACDC consists of 3D cine-MRI volumes from 100–150 patients; an image-level split puts slices (and frames) from the same patient in train and test. CVC-ClinicDB's 612 frames come from 31 colonoscopy sequences, so near-duplicate frames end up across splits. This inflates absolute numbers and may also compress differences between methods. Also, Sec. 4.3 mentions "group identifiers" used in the bootstrap, so group information appears to exist. Please clarify why the splits are not patient- or sequence-level.
  1. The quantitative metric statistics reporting has inconsistencies
  • Macro SD vs. per-dataset SDs. Each dataset is a separate training run, so per-dataset results should be roughly independent across datasets within a "matched seed," and the SD of the macro-average should be about √(Σσ²)/4. For Full CTKAN (Table 4) that gives ≈0.40, but Table 2 reports 0.15. For Static KAN it gives ≈0.48, but Table 2 reports 0.88. These are exactly the two rows that drive the headline ablation conclusion. Please explain how the macro SDs were computed.
  • Same architecture, very different variance. The paper states that Static KAN (Table 2) and U-KAN (Table 1), and Full CTKAN and CTKAN-Light, are identical architectures run with different seeds. Yet GlaS IoU SD is 1.02 vs 0.14 for Static KAN vs U-KAN, and 0.70 vs 0.16 for Full CTKAN vs CTKAN-Light (Tables 3, 4). That suggests seed-set variability large enough to matter for 0.2–0.8-point ablation gaps.
  1. The fairness of baseline comparison
  • All baselines share one generic protocol (Adam, lr 1e-4, 300 epochs, batch 4), whereas the CTKAN dynamics get a separate learning rate (1e-3) and a validation search over T, L, R_a and R_o. MedNeXt and U-Mamba were designed and tuned inside the nnU-Net pipeline, and nnU-Net itself is absent, even though it is the standard strong baseline in medical segmentation.
  • None of the KAN-based segmentation variants cited in the paper (TransUKAN, MM-UKAN++, UUEKAN) are compared, so it is unclear whether CTKAN beats the current KAN state of the art or only U-KAN. Gains are small and uneven. Against MedNeXt, CTKAN-Max is lower on CVC IoU (88.15 vs 88.31) and ACDC IoU (84.94 vs 84.98), and the macro advantage is driven mostly by BUSI (65.78 vs 62.98). Against U-KAN, the GlaS gain (+0.31 IoU) is within one SD.

Minor: 6) Clarity and framing:

  • The "thinking," "thought ticks," "neural dynamics" and "synchronization" language is anthropomorphic and overstates the mechanism, which is a weight-tied recurrence on a pooled per-sample vector with no new input across ticks. The connection to the Continuous Thought Machine is only by analogy and could be toned down.
Questions for authors
  1. Which L was used for the "CTKAN (T = 1)" ablation? What is its test performance at its validation-best L (L = 15, val 81.34 IoU)?

  2. How were macro-average SDs computed, given the apparent mismatch with per-dataset SDs (Full CTKAN 0.15 vs ≈0.40 expected; Static KAN 0.88 vs ≈0.48)? Please also explain why several Dice SDs exceed the corresponding IoU SDs.

  3. Why are splits at the image level instead of the patient (ACDC) or sequence (CVC-ClinicDB) level, given that group identifiers exist? Can you report ACDC with the official patient-level split, 3D Dice and HD95, and GlaS on the official Test A/B?

  4. How does CTKAN compare with nnU-Net and with at least one recent KAN segmentation model (TransUKAN, MM-UKAN++ or UUEKAN) under the same protocol? Were any baseline hyperparameters tuned?

  5. What happens with dataset-average gates or shuffled gates at test time (W8)?

  6. How does CTKAN compare with a non-recurrent gate generator of matched FLOPs and parameters (e.g. a deeper MLP on the same pooled statistics), and with an SE-style or dynamic-convolution gate on the KAN input?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.