acceptodds
Withdrawn from ICLR 2027

Reliability-Controlled Competitive Assignment for Conflict-Free Pseudo-Labeling in Few-Shot Multimodal Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Prototype-guided self-training expands a small labeled set by assigning unlabeled instances to class prototypes. Under few-shot supervision, however, overlapping class-specific candidate sets create an assignment ambiguity: independent per-class selection can admit the same instance with conflicting pseudo-labels. Confidence thresholds and class-aware admission rules regulate sample selection but do not, by themselves, resolve competition between classes. We propose **reliability-controlled competitive assignment**, which scores candidate assignments using a dual-constraint reliability criterion and resolves cross-class conflicts through a per-instance argmin before admission. This enforces a one-instance–one-label constraint, guaranteeing an admitted pool free of duplicate-instance label conflicts without assuming that every assigned label is correct. We integrate competitive assignment into a progressive self-training framework that combines class-balanced, confidence-guided pool expansion with momentum-smoothed pseudo-label refinement. Experiments on two few-shot multimodal classification benchmarks, **MUUFL Gulfport** and **Trento**, compare competitive assignment with independent per-class selection using matched-seed paired runs. Competitive assignment improves mean accuracy by **7.95 percentage points on MUUFL** (95% confidence interval: **[3.15, 12.75]**) and **4.79 percentage points on Trento**, although uncertainty is greater in Trento’s five-split evaluation. These findings demonstrate that resolving cross-class assignment conflicts before admission can improve few-shot multimodal classification and establish assignment consistency as a useful complement to confidence-based pseudo-label selection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Few-shot multimodal learning, pseudo-labeling, self-training, competitive assignment, conflict-free assignment, prototype learning, multimodal classification, remote sensing

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_bqdWOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses pseudo-label ambiguity in few-shot multimodal classification by introducing a competitive ownership resolution mechanism. The key insight is that independently constructed class-wise candidate pools can lead to conflicting labels for the same instance, where the final assignment depends on processing order rather than evidence. The proposed method assigns each multi-claimed instance to its most reliable class via an argmin operation before budget allocation, ensuring order-invariant and conflict-free pseudo-label pools. Experiments on MUUFL and Trento datasets demonstrate accuracy improvements with confidence intervals provided.

Critical strengths

1.The paper rigorously separates the selection decision from the ownership decision. This conceptual clarity, formalized in Proposition 1, provides a principled framework that could be adopted by other pseudo-labeling systems beyond the specific PGPL instantiation tested here.

2.The authors provide matched-seed paired experiments across multiple splits, report confidence intervals, and include extensive diagnostic analyses.

Critical weaknesses

1.The method is evaluated only within the PGPL framework on two relatively small benchmarks.

2.The competitive assignment requires computing scores for all claimants of each multi-claimed instance and performing an argmin operation. While this is tractable for the 11-class MUUFL and 6-class Trento datasets, the computational and memory implications for problems with hundreds or thousands of classes are not discussed.

3.In the figures, several connectors are misaligned, and spacing is inconsistent.

Questions for authors
  1. Can you report the computational overhead of the competitive assignment step relative to the baseline, particularly as a function of the number of classes and the proportion of multi-claimed instances?

2.The paper demonstrates the method on two datasets from the same domain (hyperspectral-LiDAR fusion). Do you have any evidence that the ownership resolution mechanism transfers to other multimodal settings (e.g., RGB-D, text-image)?

1 comment

Thank you for asking. The change reflects my reassessment of the significance of the contribution. The additional controls suggest that the gains are largely associated with class-wise admission counts and pool composition. While these results clarify how the method works, I remain unconvinced that the methodological advance is sufficiently substantial for acceptance. My revised rating reflects this concern about the contribution’s significance.

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

This paper proposes Reliability-Controlled Competitive Assignment for conflict-free pseudo-labeling in few-shot multimodal learning. The method addresses assignment ambiguity caused by overlapping class-specific candidate sets by comparing reliability scores and uniquely assigning each instance to an eligible class before budgeted admission. Experiments on MUUFL and Trento show improvements in pseudo-label purity and classification accuracy. The results are encouraging, particularly when there is substantial overlap between candidate sets.

Critical strengths
  1. The paper clearly identifies the problem of assignment ambiguity: when class-specific candidate sets overlap, the same instance can receive conflicting pseudo-labels, and its final label may depend on class processing order. This is a meaningful issue to study because arbitrary assignment decisions can introduce noise into the training pool and affect subsequent self-training.

  2. The method is simple and improves results on both datasets without changing the underlying network architecture. Competitive assignment increases pseudo-label purity and mean classification accuracy on MUUFL and Trento, making it a potentially useful modification to the existing PGPL pipeline.

Critical weaknesses
  1. The experiments lack a direct baseline that performs independent per-class admission first and then resolves duplicate assignments by retaining, for each duplicated instance, the lowest-score label among the classes that actually admitted it. Comparing this baseline with the proposed method would help clarify whether resolving assignment before budgeted admission improves pool quality and downstream accuracy over post-admission conflict removal.

  2. The assignment rule does not depend on a particular network backbone, but it is evaluated only within PGPL. The current results support a useful improvement to that pipeline; evaluation in other compatible pseudo-labeling frameworks is still needed to establish its general applicability.

  3. Some figures are difficult to read. In Figures 1 and 2, parts of the text overlap with graphical elements and lines. The text in Figure 3 is also too small. Adjusting the layout and increasing the font size would improve readability

  4. presentation and novelty. In line 78 and line 79 of page 2, the presentation of "the lowest reliability score" is confusing. The presentation of 4.1 lacks some necessary steps of the proposed few-shot multimodal learning. The novelty of the proposed method is not well presented, and the current version shows limited methodology novelty.

Questions for authors
  1. How does the proposed method compare with independent per-class admission followed by minimum-score deduplication among the classes that admitted each instance? Does resolving assignment before admission improve pool quality or downstream accuracy?

  2. To what extent do the accuracy gains come from resolving conflicting assignments versus filtering out low-quality pseudo-labeled samples?

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

The paper notices an insteresting problem in few-shot learning, independent per-class selection can admit the same instance with conflicting pseudo-labels. Built upon the existing PGPL framework, the proposed method improves pseudo-label consistency and downstream classification accuracy on two HSI–LiDAR benchmarks.

Critical strengths
  1. The paper is well-constructed.
  2. The paper acheives consistent improvement over evaluation benchmarks.
Critical weaknesses
  1. The improvement may mainly come from using fewer pseudo-labels rather than better label assignment. On MUUFL, the proposed method removes 33.2% of the pseudo-labels but only changes the labels of 1.0% of samples. Also, the class-wise selection budget is recalculated after assignment, so the two methods use different numbers of pseudo-labels even with the same ρ\rho. A comparison with the same number of pseudo-labels and a simple strategy that removes samples claimed by multiple classes would help clarify the actual benefit of the proposed assignment.
  2. The novelty of the method is limited. The reliability score, candidate filtering, and progressive self-training are all inherited from PGPL. The main contribution is simply applying an argmin operation to assign each sample to a single class before pseudo-label selection. This is a straightforward way to resolve label conflicts, and it is unclear how it differs from standard label assignment strategies commonly used in pseudo-labeling methods.
Questions for authors

Please refer to weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.