acceptodds
Under review as a conference paper at ICLR 2027

Value-Twisted Sequential Monte Carlo for Task-Aware Missing-Modality Connectome Completion

Abstract

Completing a missing modality in the field of medical imaging data could create a Prior–Utility Dilemma: samples that faithfully follow a conditional generative prior may not help prediction, while samples aggressively optimized for a classifier may no longer represent that prior. Our goal is to make diffusion-generated connectomes more useful downstream while keeping their sampling distribution explicit and auditable. We introduce Value-Twisted Sequential Monte Carlo (VT-SMC), a task-aware inference procedure for missing-connectome completion. To resolve this dilemma, VT-SMC formulates task-aware connectome completion as inference under an explicit utility-tilted generative distribution. It combines a conditional diffusion prior with frozen uncertainty-aware task feedback, then uses corrected particle inference to favor diagnostically useful connectomes without losing distributional accountability. This unified formulation balances anatomical plausibility, downstream utility, and subject-specific uncertainty, while admitting a consistent SMC estimator and a gradient-free variant. Sufficient experiments show that the method manages, rather than eliminates the fidelity–utility trade-off.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Missing-Modality Learning, Multi‑Modal Imputation, Diffusion Models, Sequential Monte‑Carlo, Brain Connectome, Uncertainty-Aware Learning, Medical Image Computing

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_zszoOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose VT-SMC, a twisted sequential MC sampler that completes missing dMRI connectomes from fMRI/sMRI while favoring completions a downstream diagnostic classifier finds useful. A label-free conditional diffusion prior is tilted by the predictive entropy of a frozen classifier ensemble. The authors show that a fresh classifier trained on the guided completions reaches 81 AUC on ABIDE II vs. 76 with unguided completions, demonstrating the utility of the method.

Critical strengths

S1) Declaring the tiled target explicitly and correcting the guided proposals is more principled than ad hoc classifier guidance.

S2) The evaluation separates utility from fidelity and uses an independent evaluator.

Critical weaknesses

W1) The submission is either unfinished, or AI-generated text was left in. Tables 5-7 are explicitly declared to be placeholders, Table 4 and Appendix C.7 say values await the "final server log," Appendix C.9 says formal experiments start after checks pass, Appendix E is an internal figure spec, etc. Additionally, Table 1 is unreadable as structured.

W2) Training-fold guidance leaks labels. C0C_0 is trained with labels on the training subjects (per Alg. 2) and then guides those same subjects' completions. This breaks the conditional independence premise of Appendix A.1. Guidance must be cross-fitted.

W3) The ablation pattern is not really credible. Table 8 shows all variants produce nearly identical outputs. Additionally, no standard error is given for the 3 seeds, 5 folds.

W4) Novelty and guarantees are thin. The sampler is twisted diffusion sequential MC sampler with an entropy potential. Prop 1 is the equal-covariance Gaussian KL identity and Prop. 2 is standard Feynman-Kac consistency.

W5) Overall, the paper is written very poorly. Beyond the unfinished parts, key symbols are undefined or defined after use (e.g., quality-tier embedding ere_r, or the operator TkT_k in Eq. 4, etc.), writing is poor and difficult to follow, and overall presentation is not very good. This text needs another human pass to get it in good shape.

Questions for authors

Q1) Are training-subject completions guided by C0C_0 models trained on those subjects' labels? IF so, does the gain survive out-of-fold C0C_0 guidance?

Q2) How does the full VT-SMC gain +5 AUC when its two ingredients alone gain <1 and Table 8 shows the variants' outputs nearly identical?

Q3) Which Table 1 entries were re-run on your splits, which are copied from papers, and which tables and figures in the submission are final?

Q4) Fidelity is validated only on the 148 subjects from the four dMRI sites. What supports generated dMRI for the 654 subjects from the other eleven sites who are also older?

Reviewer_2RohOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Value-Twisted Sequential Monte Carlo (VT-SMC) for task-aware missing-connectome completion. It combines a conditional diffusion prior, a frozen classifier-entropy potential, and importance-corrected particle sampling to improve downstream prediction while explicitly defining the resulting sampling distribution.

Critical strengths

The method clearly separates task-guided sampling from distribution correction. The evaluation includes synthetic validation, independent classifiers, and reconstruction analysis beyond downstream accuracy.

Critical weaknesses
  1. Important experimental evidence remains unfinished, and the quantitative reporting needs reconciliation. Tables 5–7 explicitly contain placeholder results, while parts of the implementation description still await final configuration logs. There are also unexplained differences between the main and supplementary reconstruction correlations.

  2. The manuscript suffers from unclear writing, poorly designed figures, and weak overall presentation, making it difficult to follow. Substantial improvements are needed to enhance clarity and readability.

Questions for authors
  1. Why do Appendix Tables 5–7 contain placeholder values rather than actual experimental results? Does this indicate that these experiments were not completed before submission? Including unverified numerical results in a submitted paper raises serious concerns about the reliability of the experimental evidence and research integrity.

  2. What is the purpose of Appendix C.5 and C.9? These sections largely describe planned validation procedures rather than completed experiments, which seems inappropriate for a submitted research paper. Have these checks actually been performed, and if so, why are their results not reported?

Reviewer_BH6COpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates the application of diffusion model in generating plausible connectomes when structural brain connectivity is missing. The designed method use a conditional diffusion model to generate plausible structure, use a classifier ensemble and entropy to prefer samples that produce confident answers, and a sequential Monte Carlo to resample particles, and correct importance weights. Finally they use the generated connectomes to train a downstream diagnostic classifier.

Critical strengths
  1. Problem is well-motivated and interesting.
  2. The use of frozen pilot uncertainty, sampling in the symmetric connectome space, and a dimension-independent proposal KL budget reflects attention to sampling correctness and numerical stability.
Critical weaknesses
  1. Major methodological weakness is that the used entropy objective does not measure ensemble agreement. Eq. 6 which is defined entropy value objective, motivated as ensemble agreement, and cross-seed consensus. However, minimizing the average entropy of individual classifiers does not encourage agreement between classifiers. This means the proposed guidance can favor samples on which individual models are confident but disagree, contrary to the paper's interpretation of the potential.

  2. Very important baselines are missing. The entire motivation is that generating task-aware structural connectomes improves classification when dMRI is missing. But the authors do not demonstrate that an expensive generative sampling pipeline is necessary to obtain these improvements. Appendix C.8 points to several appropriate controls; (1) Deterministic matrix regression (2) Latent imputation (3) Modality dropout and teacher–student distillation (4) Best-of-N selection from unguided samples. However, the main table do not report results for these baselines.

  3. The theoretical novelty is very thin. Proposition 1 is mainly a standard equal-covariance Gaussian KL calculation, and Proposition 2 invokes established SMC consistency results. These results do not establish a new theoretical result.

  4. The experimental improvement is less impressive than what the paper claims. On ABIDE II, classification AUC increases from 76.02% with unguided diffusion to 81.01% with VT-SMC, good if robust. However, reconstruction PCC decreases from 0.8645 to 0.8324, and MAE worsens from 0.0153 to 0.0173. The AUROC improvement is only 0.8072 to 0.8172.

  5. Contribution is narrow. The method seems to be mix of known components such as tilted target, shifting proposals, applying Gaussian likelihood-ratio corrections, and using SMC. The contribution is combining these with a particular entropy objective, frozen pilot uncertainty, and symmetric-connectome sampling.

Questions for authors
  1. Could the authors please justify the use of the ensemble-entropy objective?

  2. Could the compare against non-generative models as well? Or include a discussion which baselines would seem relevant.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.