acceptodds
Withdrawn from ICLR 2027

GIP-GEO: Geometric Self-evaluation for Few-shot Class-incremental Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-shot class-incremental learning requires learning novel classes from few labeled examples while retaining previously learned classes. Class-nameconditioned diffusion supports training-free adaptation, but its descriptions may not reflect the observed support images, and its generations are typically used without an explicit alignment check. GIP-Geo addresses both limitations. Grounded Iterative Prompting conditions the language model on support images and nearby base-class names, producing candidate descriptions of observed, discriminative attributes. Geometric self-evaluation then scores agreement among each candidate’s text, real-support prototype, and generated prototype via alignment-regularized volume (ARV), which combines a kernel Gram determinant with pairwise dispersion to avoid a degenerate zero score. The same score selects one candidate and its fusion weight with real supports. On CIFAR-100, miniImageNet, and CUB-200, exhaustive GIP-Geo improves average accuracy over the controlled CD-FSCIL baseline without incremental network training; an optional score proxy reduces generation cost while CLIP and diffusion remain frozen.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Few-shot Class-incremental Learning, Training-free

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_WfR4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces GIP to ground candidate descriptions in support images and retrieved base class contrasts. It leverages the Alignment-regularized volume (ARV) to measure the agreement among each candidate’s text, real support prototype, and generated prototype for candidate ranking, so as to achieve training-free few-shot class incremental learning. However, the manuscript is loosely written with poorly formatted figures and obvious AI-generated artifacts. The technical contribution is limited to an improved alignment scheme, making the work incremental and lacking novelty.

Critical strengths

A novel alignment strategy is proposed to realize training-free few-shot class incremental learning.

Critical weaknesses
  1. The manuscript exhibits strong signs of AI generation. The introduction suffers from confusing logic and obscure wording, and its brevity hinders clear understanding of the core contributions.
  2. Figure 1 is poorly designed, filled with excessive text and drawn casually. The same issue applies to Figures 3 and 4.
  3. In the experimental section, only CD-FSCIL (2025) is adopted as a recent baseline, while all other compared methods are outdated. Comparisons against more recent state-of-the-art (SOTA) approaches are missing.
  4. The proposed method merely modifies the alignment mechanism. The work is simplistic and lacks sufficient innovation.
Questions for authors

None.

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

This paper proposes GIP-Geo for few-shot class-incremental learning. The method uses GPT-4o to generate class descriptions based on support images and related base-class names. A diffusion model then generates additional samples conditioned on these descriptions. The authors introduce an Alignment-Regularized Volume (ARV) score to evaluate the consistency among text, real-support, and generated-image features. This score is used to select generated samples and determine prototype fusion weights. Experiments on CIFAR-100, miniImageNet, and CUB-200 show improvements over the CD-FSCIL baseline. Additional experiments investigate the contributions of different components and the efficiency of the proposed method.

Critical strengths
  1. The idea of using support images to guide the generation of additional training samples is meaningful for few-shot class-incremental learning.
  2. The overall method achieves consistent accuracy improvements over CD-FSCIL across three datasets. The additional experiments with multiple support seeds also provide useful evidence for the performance gains.
Critical weaknesses
  1. The benefit of the proposed ARV score is not sufficiently demonstrated. Table 5 shows that simpler metrics, such as dispersion-only and cosine disagreement, achieve comparable or even better accuracy than ARV. Since ARV is one of the main technical contributions, I find it difficult to justify the necessity of this more complicated geometric formulation. Moreover, the idea of evaluating text, real, and generated features using kernelized volume has already been explored in VT-FSL. The methodological novelty is therefore somewhat limited.

  2. The experimental validation is incomplete. Several ablation results come from different experimental settings and cannot be directly compared. The paper also reports unexplained differences between some results and missing records for previous experiments. More importantly, there is no matched support-only baseline or direct comparison with VT-FSL adapted to this setting. These missing controls make it difficult to determine how much of the improvement actually comes from the proposed contributions.

Questions for authors

See Weaknesses.

Reviewer_6uTaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The manuscript introduces GIP-Geo, a training-free adaptation framework for few-shot class-incremental learning (FSCIL). The method targets two primary bottlenecks in class-name-conditioned diffusion generation: the misalignment between broad class names and observed support images, and the lack of explicit alignment validation for generated exemplars.

To resolve these issues, the framework proposes:

  1. Grounded Iterative Prompting (GIP): Utilizes an MLLM (GPT-4o) conditioned on support images and text embeddings of nearest base-class neighbors to generate candidate descriptive attributes.

  2. Alignment-Regularized Volume (ARV): A geometric self-evaluation score combining a kernel Gram determinant with pairwise dispersion to score text-support-generation triplets, enabling candidate selection and prototype fusion without query supervision.

Critical strengths
  1. Creative Problem Formulation: The shift from unguided text-to-image generation to support-grounded, context-aware condition generation is a conceptually sound direction for mitigating domain drift in generative FSCIL.

  2. Thoughtful Mathematical Refinement of Geometric Scores: The introduction of pairwise dispersion into the volume-based metric (ARV\mathrm{ARV}) successfully resolves the degenerate zero-score ambiguity that occurs when input vectors coincide, presenting an interesting geometric formulation.

Critical weaknesses
  1. Unprofessional Meta-Commentary and Transparency Disclaimers: The manuscript reads more like an internal engineering log or an AI-generated technical report than a polished scientific paper. It frequently breaks the fourth wall to admit a lack of reproducibility: Example: "The archived ranking audit lacks a run-specific classorder record and initial proxy checkpoint, so its exact sequence cannot be reconstructed from the report alone". Example: "archived benchmark records do not identify a dated model snapshot. We therefore cannot verify the exact GPT-4o version used in each historical run". Admitting these limitations is honest, but publishing them as an excuse for unverified checkpoints and missing metadata fails standard peer-review benchmarks.

  2. Unstable and Exploratory Modules Presented as Core Contributions: The "classification correction" (reliability penalty) is shown to introduce severe instability—suppressing novel class accuracy from 65.92% down to 18.33% on certain support seeds due to hard class deletion. While the authors demote this to an "exploratory extension," its inclusion clutters the narrative and indicates that the framework's hyperparameter tuning (λconf=4.89\lambda_{conf} = 4.89) lacks cross-seed robustness.

  3. Inconsistent Component Ablations and Unmatched Controls: Several core component tables (e.g., Table 3) rely on "replacement measurements from an earlier penalty-enabled setting," meaning they do not isolate component performance under the final correction-off core policy. This lack of clean, isolated ablations weakens the empirical claims.

  4. Extensive Reliance on AI-Assisted Disclosures: The AI Use Statement explicitly notes that Codex assisted with the “formulation and revision of mathematical arguments” and “interpretation of reported results”. While leveraging tools is acceptable, key mathematical justifications (such as the convergence properties of the prototype perturbation bounds in Appendix B) carry heavy reliance on automated generation without apparent independent verification.

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.