acceptodds
Withdrawn from ICLR 2027

From Forward Prediction to Self-Corrective World Modeling for MRI Contrast Enhancement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion models offer strong visual quality for video generation and are promising for MRI enhancement world modeling. However, existing diffusion-based MRI enhancement world models remain passive forward simulators: they can predict future states, but cannot recognize and correct emerging kinetic errors. For the first time, we propose ReKinetics, a self-corrective world model that extends diffusion-based world modeling from forward prediction to explicit error correction. For self-correction action supervision, we introduce Corrective Credit Assignment Learning (CCAL), which shifts supervision from trajectory-level error fitting to action-level corrective credit assignment, turning latent kinetic failures into executable where–when–how action targets. For action execution, we introduce Factorized Corrective Transition Learning (FCTL), which replaces the shared denoising transition with a factorized corrective transition parameterized by state-dependent activation, spatial support, and action-specific residual transformation, decoupling heterogeneous corrective effects within diffusion. For action feedback, we introduce Reliability-Guided Corrective Optimization (RGCO), which shifts optimization from prediction-error-driven learning to correction-outcome-driven learning by converting reliability changes into trajectory-specific optimization weights. Together, these mechanisms form a closed self-corrective learning loop. Across Hand DCE-MRI and Duke Breast, ReKinetics achieves strong frame and kinetic performance. On the Hand benchmark, it reduces late-vs.-early ROI kinetic error by 42.1%, whereas competing methods accumulate 23.7%–44.7% error, confirming that explicit kinetic correction substantially benefits dynamic contrast-enhancement modeling.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Self-Corrective World Models, Diffusion Models, Corrective Policy Learning, Kinetic Modeling, DCE-MRI

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

This paper presents ReKinetics, a diffusion-based framework for synthesizing dynamic contrast-enhanced MRI from non-enhanced inputs and acquisition times. The method learns corrective interventions through three components: CCAL constructs structured failure cards from training-time prediction discrepancies to supervise action routing; FCTL injects learned corrective residuals into denoising through state-dependent activation and spatial support; and RGCO reweights subsequent training according to changes in reconstruction-based reliability. Diagnostic supervision is distilled into internal routing and kinetic-state modules, enabling inference without an external diagnostic model or reference trajectory.

Critical strengths
  • Kinetic evaluation: The paper evaluates regional enhancement trajectories alongside frame fidelity. Figure 3 and the target-referenced crSSIM metric help assess whether generated sequences reproduce enhancement changes beyond plausible individual images.
  • Explicit corrective mechanism: Sections 3.1–3.2 connect structured diagnostic supervision to action activation, spatial support, and learned residual transformations. This makes corrective execution concrete and allows inference without the external diagnostic model used during training.
Critical weaknesses
  1. The paper proposes too many modules and losses and lacks sufficient ablation experiments to demonstrate the effectiveness of each component design. The paper introduces seven losses and three modules, but not every loss or module has a corresponding ablation study.
  2. Why not implement the self-correction idea using a large model? Implementing a small routing component inside a diffusion model makes it difficult to truly "understand" what the error is. Even if improved generation is observed, describing the effect as "self-correction" is somewhat forced and overstates the model's capability.
  3. Taking Figure 3(a) as an example, the cumulative error metric also decreases at points not labeled as self-correction. Does this indicate that Figure 3(a) alone cannot confirm that the effect is due to self-correction?
  4. Table 1 shows inconsistent trends across the two datasets; perhaps more datasets are needed to validate the framework's effectiveness.
Questions for authors

See Weaknesses

Reviewer_5HRGOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper predicts dynamic contrast-enhanced MRI (DCE-MRI) frames from one non-enhanced slice and an acquisition time. ReKinetics adds three parts to ControlNet latent diffusion.

  1. Corrective Credit Assignment Learning (CCAL) uses fixed rules and GPT-5.5 to turn errors against the reference into a "Failure Card". The card names the failure, its location and which of 14 corrective actions to apply.

  2. Factorized Corrective Transition Learning (FCTL) adds the routed actions as gated residuals to the noise prediction.

  3. Reliability-Guided Corrective Optimization (RGCO) upweights trajectories, slices followed over time, that keep high error.

At inference, a preliminary rollout feeds a router that gates the actions in a refined rollout. Data are a private hand cohort (34 subjects, 7 in test) and Duke Breast. Table 1 reports the best PSNR on both sets. Error-focused refinement is a reasonable idea. But three Table 3 rows lie on the line between two other rows, so the component gains do not appear to be measured.

Critical strengths
  1. A matched Action ON versus Action OFF control on the full test set. Appendix C.5 runs all 406 test trajectories twice with the same input, checkpoint, noise and seed. Action OFF zeros only the routed corrective residual. This is the right design for the inference-time claim. Table 11 reports wins in 290 of 406 trajectories for region-of-interest (ROI) mean absolute error (MAE). This experiment could carry the paper once effect sizes are added.

  2. Reproducible supervision rules and a clean hand split. Appendix A.1 gives the diagnostic thresholds and filter sizes. Table 7 lists all 14 actions. Table 10 gives the seed, update count, batch size and hardware. Table 9 splits the hand data by patient. Failure Cards use only the 20 training subjects (page 19, lines 1022 to 1023).

Critical weaknesses

1. Three rows in Table 3 fall on one straight line. So the gains of the components do not look like real measurements.

Table 3 is the paper’s evidence that each component improves the model (page 9, lines 464 to 472). I believe the FCTL, CCAL and KAN rows were calculated, not measured. Here is why.

Take two rows of Table 3: the capacity-control row and the full-model row. The FCTL, CCAL and KAN rows sit exactly 0.20, 0.55 and 0.90 of the way between them. This is true for all 18 numbers in these three rows, at the precision printed in the paper.

An example: crSSIM measures how well the predicted enhancement matches the true enhancement (Eq. 28). From the capacity-control row to the full model, crSSIM rises from 0.4401 to 0.8161. The total gain is 0.3760. Add 0.20, 0.55 and 0.90 of this gain to 0.4401. The results are 0.5153, 0.6469 and 0.7785. These are exactly the numbers printed in the table.

As a result, the CCAL step and the KAN step give the same gain in five of the six columns. For example, both add 2.2846 dB of PSNR.

This pattern appears nowhere else in Tables 1 to 3. Models trained separately do not fall on one straight line at simple fractions. Averaging the outputs of two models would not produce this either, because PSNR and SSIM do not change linearly with the outputs.

The KAN row also appears again in Table 2, Panel B, where it is compared with a multilayer perceptron (MLP) of the same size. If these rows were not measured, then no gain in Table 3 and no KAN result in Table 2 is supported by an experiment.

Please explain how these rows were produced. If they came from training runs, please provide the training logs, the checkpoint names and the results for each trajectory. If they did not, please replace them with measured results. This problem is serious enough to reject the paper, because Table 3 is the only experiment that adds the components one at a time.

2. The main kinetic result has no effect size, and the Duke baselines are copied from another paper.

The abstract says that the region-of-interest (ROI) error drops by 42.1% from early to late time points. It also says the error rises for all 5 competing methods (page 1, lines 030 to 032). This percentage does not show that the model corrects its errors.

In Figure 3(a), ReKinetics has the highest error at the start and the lowest at the end. Its first value is about 0.24. The other methods start at 0.14 to 0.19 (read from the plot). A high starting error alone makes the drop look large. The paper also does not define “early” and “late”.

Table 11 is the direct test of correction. But it only counts how often the model wins. It does not say by how much. Also, its 406 trajectories are slices from only 7 test patients.

On the Duke dataset, the six baseline rows (page 7, lines 345 to 349) are copied. All 30 numbers are the same as in Table 1 of MRI CEKWorld (Kong et al., CVPR 2026). The paper says it follows published protocols (page 6, lines 311 to 312). It never says the numbers come from another paper. Neither paper reports how Duke was split or how many test cases there are.

The CCNet row has a further problem. The CEKWorld authors say their CCNet training did not converge. This row has an LPIPS of 0.3128. In the CCNet paper itself, simply using the image before contrast already scores better (0.199 to 0.223; Osuala et al., MICCAI 2024).

The Duke cSSIM metric (Eq. 29) gives a perfect score of 1 to an output that never changes over time. But Section 4.3 still uses it as evidence of good behavior over time (page 7, lines 353 to 355).

The authors can answer these points in the rebuttal:

  • Report how much the error changes when the corrective actions are on versus off, for each patient. Include the cases that get worse.
  • Report the Duke split. Then re-test at least ControlNet and CEKWorld on that split.
  • Replace the CCNet row with a run that converged, or remove it.
  • On both datasets, add a simple baseline that uses the input image as the prediction.
  • Report the cSSIM of the real image sequences, for comparison.
Questions for authors
  1. Which router input carries an error signal at inference? There, the kinetic deviation is the preliminary rollout minus the non-enhanced input (page 5, lines 256 to 258). That is predicted enhancement, not error. The severity input is zero.

  2. How were the evaluation ROIs in Eq. 30 chosen? Exams have 12 to 26 acquisitions. How do they map to the 19 points of Figure 3(a)? How many trajectories does each point average, and is it weighted by patient or by slice?

  3. How many epochs and what learning rate trained the ControlNet backbone row (20.753 dB) and the other five hand baselines? Table 10 lists a learning rate of 2 x 10^-7 for ReKinetics. Is that value correct?

Reviewer_Ssr7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ReKinetics, a diffusion-based method for predicting contrast-enhancement trajectories from non-enhanced MRI and acquisition times. During training, diagnostic tools and a VLM produce structured failure cards that supervise where, when, and how corrections are applied. Learned residual operators carry out these corrections within the denoising process, while reconstruction errors and their changes across training rounds guide trajectory-specific weighting. The authors evaluate the method on hand and breast DCE-MRI datasets, examining image quality, enhancement dynamics, and the contributions of individual components.

Critical strengths
  1. I appreciate the focus on enhancement curves and temporal behavior, which are central to DCE-MRI and deserve attention alongside image quality.
  2. The experimental design includes relevant controls, particularly the action on/off tests and capacity-matched KAN/MLP comparison, to examine the proposed mechanisms.
  3. Using the VLM only during training is a practical design choice. The learned correction mechanism operates without external VLM calls at inference.
Critical weaknesses
  1. I am flagging a potential research-integrity concern in Table 3. Table 3 raises concerns about the reliability of the ablation results. Across all six metrics, the FCTL, CCAL, and KAN variants match “linear interpolations" between the monolithic-residual baseline and the full model at weights of 0.20, 0.55, and 0.90, respectively, within rounding precision. Correlated metrics may improve together, but they are not generally expected to follow identical interpolation fractions across distinct ablations. This pattern requires clarification before these results can reliably support the claimed contributions of individual components.
  2. I find the method more convincingly described as a self-corrective diffusion model. I recognize that “world model” has several meanings, but temporal conditioning and diffusion generation alone do not establish this claim. I have a similar concern about the cited MRI CEKWorld. Here, the actions appear to correct generated images internally, rather than represent interventions on the underlying enhancement process. What modeling capability does this formulation offer beyond conditional generation and iterative refinement? Also, deviation from the non-enhanced baseline may reflect genuine enhancement rather than prediction error.
  3. The method clearly involves substantial engineering, but the necessity of its many auxiliary losses and handcrafted priors remains unclear. I cannot tell how much improvement comes from the correction mechanism versus additional supervision. The shared interpolation pattern in the ablation results further complicates their interpretation as evidence for individual components.
  4. The action categories are predefined, and the diagnostic tools already identify several failure types. It remains unclear how much the VLM contributes beyond these rules. A comparison using rule-generated Failure Cards alone would help answer this.
  5. The hand test set includes only seven subjects, making patient-level variability important. Duke cSSIM measures similarity between predicted frames and could favor overly smooth trajectories. Patient-level uncertainty and kinetic metrics directly comparing predictions with ground truth would strengthen the evidence.
Questions for authors
  1. How were the ablation results in Table 3 obtained, and what explains the shared interpolation weights of 0.20, 0.55, and 0.90 across all six metrics? Please provide per-case results, evaluation scripts, and corresponding experiment records to verify the separately evaluated variants.
  2. How does performance change when failure cards are generated by the diagnostic tools alone, without the VLM?
  3. How does the inference-time error proxy distinguish genuine enhancement from prediction errors, and how are states and corrections updated during inference?
  4. Would a simpler correction model achieve comparable results with the same auxiliary losses and training budget?
  5. Can you report patient-level uncertainty and kinetic accuracy measured directly against ground-truth trajectories?
1 comment
AuthorsOpenReview ↗

We thank the reviewer for the careful assessment. We corrected Table 3 and provided checkpoints for the ablations. The new controls show gains from VLM-composed supervision and factorized corrective execution under matched comparisons. We also report results for all seven hand test patients and clarify how the router updates corrections during inference without ground truth or a VLM.

1. Answer:

Thank you for identifying this issue, and we apologize for the oversight. During manuscript assembly, an obsolete intermediate version of Table 3 was inadvertently retained. The identical interpolation fractions were a table-preparation artifact and did not represent independent experimental measurements. We have replaced Table 3 with the corrected results. We also provide the corresponding checkpoints. The corrected results support our conclusion.

VariantPSNR ↑SSIM ↑LPIPS ↓rMSE ↓crSSIM ↑ROI intensity MAE ↓
ControlNet backbone20.75300.5241000.123800127.06990.4325000.132000
+ Capacity-only monolithic residual20.75460.5242660.123767127.04410.4326600.131946
+ Factorized Corrective Transition Learning20.75450.5241090.123781127.04890.4325450.131954
+ Corrective Credit Assignment Learning27.36260.8791460.10340377.93600.8171330.048897
+ KAN-Based Kinetic-State Routing27.35780.8787270.10356477.84890.8146220.048882
+ Reliability-Guided Corrective Optimization27.35940.8801000.10340077.87300.8161000.049000

2. Answer:

We replaced VLM-composed Failure Cards with a rule-only version:

Failure Card constructionPSNR ↑SSIM ↑LPIPS ↓rMSE ↓crSSIM ↑ROI MAE ↓
Rule-only26.61780.85730.107985.46210.76840.0596
Training-time VLM27.35940.88010.103477.87300.81610.0490
The VLM improves PSNR by 0.7416 dB and crSSIM by 0.0477, and reduces ROI MAE by 0.0106. Thus, its benefit comes from combining diagnostic findings into case-specific corrective supervision, beyond a fixed rule mapping.

3. Answer:

The inference-time deviation from the non-enhanced image describes predicted enhancement; it does not directly measure error or trigger correction on its own. The router combines it with acquisition timing, regional intensity trajectories, and latent features to select actions learned from target-referenced Failure Cards during training. At inference, ReKinetics extracts kinetic context from a preliminary trajectory and keeps it fixed during refinement. At each denoising step, the router uses the current latent state to update action strengths and spatial support. Those actions change the latent state, which informs the next step. No ground truth or VLM is used at inference.

4. Answer:

To separate factorized execution from supervision and auxiliary regularization, we trained a parameter-matched, fully supervised monolithic residual with the full supervision and training budget. We also removed the three temporal and spatial auxiliary losses jointly while retaining the complete ReKinetics architecture. The correction mechanism makes a measurable contribution. Moreover, turning the learned actions on improves ROI MAE in 71.4% of trajectories. Thus, the simpler model does not match ReKinetics.

Matched architecture and auxiliary-loss ablations.

VariantPSNR ↑SSIM ↑LPIPS ↓rMSE ↓crSSIM ↑ROI MAE ↓
Fully supervised monolithic residual (parameter-matched)26.78240.85870.106883.41760.77490.0587
ReKinetics w/o temporal/spatial auxiliaries27.11480.87160.102980.26350.79570.0538
Full ReKinetics27.35940.88010.103477.87300.81610.0490

5. Answer:

We report below the patient-level results from the same evaluation used for the main table. We computed all five metrics separately for each patient, averaging that patient’s slices and acquisitions first. We calculated a two-sided 95% Student’s tt interval (df=6df=6). As shown in the table below, performance varies across patients, with Cases 24 and 29 presenting the greatest difficulty on several metrics.

Patient or summaryPSNR ↑SSIM ↑LPIPS ↓rMSE ↓crSSIM ↑
Case 229.48670.90970.087260.18420.8658
Case 927.91830.88740.100971.62350.8226
Case 1326.73590.87590.108680.24910.8017
Case 1430.10420.91620.082855.37860.8834
Case 2424.06380.83780.1274104.71680.7369
Case 2923.71860.82360.1357112.93540.7132
Case 3029.48830.91010.081260.02340.8891
Patient-macro mean27.35940.88010.103477.87300.8161
95% t interval[24.9298, 29.7890][0.8461, 0.9141][0.0833, 0.1235][56.7422, 99.0038][0.7513, 0.8809]

Comments on the paper

2 comments

Thank you for the response. However, it does not resolve my concern about Table 3.

The AI Use Statement in the manuscript says that "the authors reviewed and verified all AI-assisted content." The rebuttal now describes the problematic entries in Table 3 as placeholders left in by accident. These two statements are difficult to reconcile: either AI was used beyond what was disclosed and its output was not verified, or unverified numbers were reported as experimental results. In either case, the authors are responsible for every value in the submitted tables.

Because Table 3 supports central claims, neither an explanation nor a public checkpoint evaluated on a private dataset is sufficient. Before re-evaluating, I would need:

  1. Per-case evaluation results underlying Table 3 or the corresponding experiment records (logs, configurations, timestamps).
  2. The evaluation scripts.

Without these, I cannot rely on the quantitative results, and I will maintain my current rating. If the authors provide this evidence, I am willing to raise my rating to at least weak reject, in line with my earlier assessment of the paper's overall quality, and I will then go through the responses carefully before settling on a final score.

Thank you for carefully reviewing our manuscript and identifying the errors in Table 3. We sincerely apologize for our failures in reporting and verifying the results, and for the additional burden this has placed on you. We take full responsibility. We understand that these errors have undermined your confidence in the reliability of the reported results, and we respect your decision to maintain your current assessment until sufficient supporting evidence is available. We also appreciate your willingness to reconsider our work once we provide additional experimental results and verification materials. Thank you for your time and effort.

The table below reports the filesystem modification timestamps of the six checkpoint files previously shared, in UTC−12:00.

Shared checkpointFile-modification timestamp
01: ControlNet Backbone2026-08-01 02:07:02
02: Monolithic Residual Capacity Control2026-08-01 15:26:54
03: Factorized Corrective Transition Learning2026-08-12 17:31:48
04: Corrective Credit Assignment Learning2026-08-13 06:59:14
05: KAN-Based Kinetic-State Routing2026-08-14 17:13:48
06: Reliability-Guided Corrective Optimization2026-08-15 10:33:06

The component-wise parameter analysis gives the following observations:

ComparisonCheckpoint-level observation
Shared initialization versus trained variantsThe frozen diffusion U-Net, VAE, and text encoder remain unchanged; ControlNet parameters differ.
Backbone versus monolithic/FCTL variantsThe corresponding corrective branches contain non-identical weights. Their active parameter counts are closely matched: 135,668 for the monolithic branch and 135,352 for the factorized bank.
FCTL versus CCALDifferences occur in both the internal router and corrective bank; the router’s relative L2 parameter difference is approximately 11.93%.
CCAL versus KANDifferences include the kinetic-state heads; the KAN router’s relative L2 difference is approximately 8.25%.
KAN versus RGCOExisting ControlNet, routing, and corrective parameters differ; RGCO introduces no additional inference-time module.

Relative L2 differences are calculated as ∥θB−θA∥2/∥θA∥2\|\theta_B-\theta_A\|_2/\|\theta_A\|_2 within each parameter group. The checkpoint analysis supports consistency between the released model states and the intended ablation design. The variants preserve the shared frozen image-generation backbone while differing in the intended trainable components: corrective operators, the internal router, and kinetic-state heads. The CCAL variant differs in both routing and corrective parameters; KAN adds differences in kinetic-state heads; RGCO retains the same inference architecture while changing existing weights. Thus, the released variants exhibit the expected component-specific parameter differences while preserving the common frozen representation.

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

Related papers

Discussion (0)

Sign in to comment.