acceptodds
Withdrawn from ICLR 2027

Mass-Preserving Visual–Text Fusion for Training-Free Few-Shot Class-Incremental Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Training-free Few-shot Class-incremental Learning (FSCIL) must use scarce support images without confusing class-specific evidence with probability reallocation between base and novel classes. This distinction matters because a globally normalized interface can change both the ranking within the novel group and the competition between novel and base classes at the same time. Direct support-image fusion therefore makes the source of an endpoint change difficult to identify. We introduce a Mass-preserving Visual–text Fusion (MP-VTF) correction at the Bilevel Modality Calibration (BiMC) interface to separate these effects. MP-VTF uses labeled support images during frozen inference, without updating model parameters or accessing query labels. It applies a query-wise log-sum-exp translation that retains the fused within-novel ordering while restoring the text branch’s original novel-group probability mass. Across CIFAR100, CUB-200-2011, and miniImageNet, Naive fusion increased the mean novel-group probability mass in all 78 session–seed records, whereas MP-VTF restored it to numerical precision. MP-VTF produced small positive novel-class changes, nearly unchanged overall accuracy, and a small base-class cost. The miniImageNet endpoint direction is reported descriptively because its visual weight was selected with label-dependent metrics. These results establish mass preservation as a testable interface-control principle for separating within-group evidence updates from group-level probability allocation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Few-Shot Class-Incremental Learning, Training-Free, Representation Learning

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

This paper focuses on the training-free few-shot class-incremental learning at the Bi-level Modality Calibration (BiMC) interface. This study finds that the within-novel ranking and the probability mass assigned to the novel groups are impacted by the log-sum-exp (LSE) of the fused logit drifts. The authors conduct the experiments on CIFAR100, CUB200, and miniImageNet, and the results show that the fusion lowers the performance, while MP-VTF stays close to BiMC with slightly higher novel accuracy.

Critical strengths
  1. The method is training-free with two invariances.

  2. The comparison between Naive fusion and MP-VTF is fair.

  3. The authors explicitly discuss the limits of their approach.

Critical weaknesses
  1. The novelty is very limited. The contribution is incremental compared to existing studies.
  2. The improvement is negligible compared to the baseline. For example, Avg by +0.023 / +0.002 / +0.016 and Final by +0.103 / +0.013 / +0.020 (CIFAR100 / CUB200 / miniImageNet), while the standard deviations are 0.1 - 0.4. These differences are more like noise, and a comprehensive test with three seeds is impossible.
  3. This paper states it can hold the novel-group mass fixed, but does not show that fixed mass is desirable. Appendix B shows that Visual 1NN helps and hurts nearly the same (28 vs. 27 on CUB200, seed 2), while mass preservation itself does not guarantee semantically correct reranking. Thus, the method appears to mainly mitigate Naive fusion’s accuracy loss, leaving its substantive contribution unclear beyond correcting a side effect of the baseline.
  4. The hyperparameter selection and sensitivity are not properly evaluated. For example, the visual weight w is applied the same to all datasets, and there is no sensitivity analysis of it, as well as experimental evaluation across different values.
  5. The experiments should also evaluate the performance across other backbones, training-free FSCIL methods, and class-split protocols.
Questions for authors
  1. How does MP-VTF compare with Naive fusion methods that use a tuned weight, a learned or fixed base/novel bias, or a global temperature, under the same settings?

  2. What are the results under different w values?

  3. What are the key novelty and technical contributions of this work rather than the incremental efforts?

Reviewer_6o9cOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Mass-Preserving Visual–Text Fusion (MP-VTF) for training-free few-shot class-incremental learning. It addresses the issue that directly incorporating visual support evidence can unintentionally alter the probability allocation between base and novel classes. MP-VTF applies a query-wise log-sum-exp correction to preserve the original novel-group probability mass while retaining visual–text fusion benefits within novel classes. Experiments on CIFAR100, CUB-200-2011, and miniImageNet demonstrate accurate mass preservation, modest improvements in novel-class accuracy, and nearly unchanged overall accuracy compared to BiMC.

Critical strengths

The paper distinguishes within-novel redistribution from between-group probability allocation. The proposed LSE correction enables control over the novel-group mass without requiring updates to model parameters or alterations to other inference modules of BiMC. Across three benchmarks, it maintained overall accuracy close to BiMC and yielded higher dataset-level mean novel accuracy, providing a simple interface control for training-free FSCIL.

Critical weaknesses
  1. Limited technical novelty: The proposed LSE correction is mathematically straightforward and closely resembles standard group-wise probability normalization. Its novelty beyond existing calibration techniques is not sufficiently established. Please see previous works BiC [1] and SS-IL [2], which have already investigated prediction bias and calibration between old and new classes in incremental learning.
  2. Insufficient justification for mass preservation: The paper assumes that the original Text-CMNN novel-group probability mass should be preserved, but provides limited evidence supporting this assumption.
  3. Marginal performance improvements: The overall accuracy gains over BiMC are very small, particularly on CUB200.
  4. Limited experimental validation: The experiments use only three random seeds and a single visual fusion weight. More comparisons with alternative calibration strategies and independent hyperparameter selection are needed to establish the robustness of the method.

[1]Wu Y, Chen Y, Wang L, et al. Large scale incremental learning[C]//2019 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2019: 374-382.

[2] Ahn H, Kwak J, Lim S, et al. Ss-il: Separated softmax for incremental learning[C]//2021 IEEE/CVF international conference on computer vision (ICCV). IEEE, 2021: 824-833.

Questions for authors
  1. Why should the original Text-CMNN novel-group probability mass be considered the correct reference?
  2. What is the technical distinction between MP-VTF and standard group-wise conditional normalization?
  3. Can the authors establish statistically reliable improvements under an independent hyperparameter-selection protocol?
Reviewer_F9MPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a specific inference-time pathology at the BiMC FSCIL interface. It directly fuses Visual 1NN support-image scores into Text-CMNN novel logits not only reranks novel classes, but also shifts the aggregate probability mass assigned to the novel group. Experiments on CIFAR100, CUB200, and miniImageNet demonstrate the effectiveness of the method.

Critical strengths
  1. The entanglement of within-group reranking with cross-group probability allocation under a global softmax is interesting, and the Naive-vs-MP-VTF contrast isolates it cleanly.

  2. The paper is careful to distinguish "Naive vs. MP-VTF" (isolates the correction at fixed fused input) from "MP-VTF vs. BiMC".

Critical weaknesses
  1. The observation that a group-internal additive shift redistributes probability between groups under a shared softmax is the same mechanics underlying separated softmax , grouped softmax, and temperature scaling. What remains is the choice of anchor: preserving the Text-CMNN group mass. The paper never argues why the text branch's allocation is the correct reference point.

  2. The empirical gains over BiMC are negligible. The headline deltas over the BiMC baseline are small. These are far inside noise, and no significance testing is offered.

  3. The abstract and conclusion present mass preservation as a general testable principle, but everything is demonstrated at one interface (BiMC), one backbone (ViT-B/16). For a paper whose contribution is a principle, at least a second interface or a ww -sensitivity study is needed.

Questions for authors

Please see the weaknesses above.

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

Related papers

Discussion (0)

Sign in to comment.