acceptodds
Under review as a conference paper at ICLR 2027

InfoSculpt: Rebalancing Discriminative and Discoverable Information for Generalized Category Discovery

Abstract

Generalized Category Discovery (GCD) poses a definitional tension: one representation space must simultaneously serve discrimination of known classes and discovery of unseen ones. We present evidence that a root cause of current GCD methods' failures on novel classes lies in unconstrained information allocation within the representation itself. Spectral analysis of a representative trained GCD model reveals two opposite pathologies coexisting in its novel-class features—over-compression into narrow manifolds, and under-compression that buries category cues in instance noise—while downstream remedies (debiasing heads, pseudo-label repair, decoupled branches) act at the classifier output and leave a residual gap. We then audit the allocation causally: through a conditional mutual information (CMI) variational bound, the trade-off is made an explicit knob, and under a matched-seed significance protocol, turning it toward discoverability improves New-class accuracy on 14/14 method–dataset pairs (+1.3 to +13.2 pp), significant in 13/13 held-out tests (Holm-adjusted p ≤ .008) while old-class accuracy is preserved on average, and it outperforms single-term and static-balancing alternatives. Integrated into five SOTA methods across seven benchmarks, InfoSculpt raises All accuracy by +2.3 pp on average (up to +17.2 pp) and ranks best among all compared methods on five benchmarks, at 1.18× end-to-end cost with zero inference overhead.

Primary area:
interpretability and explainable AI
Keywords:
Generalized Category Discovery (GCD), latent space sculpting, conditional mutual information, information bottleneck, Nash‑MTL, representation learning, open‑world recognition

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_uG38OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes InfoSculpt, a post-training refinement framework for generalized category discovery (GCD), addressing the imbalance between discriminative and discoverable information in feature representations. It introduces category-level and instance-level conditional mutual information objectives to improve class discrimination and preserve novel-category structure, with Nash bargaining dynamically balancing these objectives. Experiments across seven benchmarks and five GCD methods demonstrate improved recognition performance, achieving an average overall accuracy gain of 2.3 percentage points with modest training overhead and no additional inference cost.

Critical strengths
  1. The paper identifies information allocation imbalance as a key limitation in GCD and introduces a unified information-theoretic framework with adaptive Nash bargaining. This provides a meaningful perspective on improving both known-class discrimination and novel-category discovery.
  2. InfoSculpt consistently improves performance across seven benchmarks and five GCD methods. The extensive ablation studies and statistical significance tests further support its effectiveness, while the absence of additional inference overhead strengthens its practical value.
Critical weaknesses
  1. The proposed category-level and instance-level CMI objectives build upon existing information-theoretic regularization, while Nash-MTL is an established optimization strategy. The authors should better clarify the fundamental novelty beyond combining these techniques.
  2. Although the paper emphasizes improving novel-category discovery, Tables 1–2 show that New accuracy sometimes decreases after applying InfoSculpt, particularly for stronger baselines such as ProtoGCD and AptGCD. These inconsistencies raise concerns about the effectiveness and reliability of the proposed information-balancing mechanism across different methods.
Questions for authors
  1. Clarify the fundamental differences between InfoSculpt and existing information-theoretic regularization methods, particularly how the dual-level CMI formulation contributes beyond conventional representation learning objectives and Nash-MTL.
  2. Explain why InfoSculpt sometimes reduces novel-class accuracy for stronger baselines in Tables 1–2, despite its goal of improving category discovery. Clarify the differences between these results and the consistent improvements reported under the discovery-oriented preference setting in Section 4.2.
2 comments
AuthorsOpenReview ↗

We thank Reviewer uG38 for the constructive feedback.

Q1. Contribution and positioning

Our contribution is to turn the representation conflict between known-class discrimination and novel-class discovery in GCD into an explicitly controllable information-allocation problem. Spectral analysis reveals over-compression and under-compression in the novel-class representations of the same model, motivating an explicit choice of which variations to compress under which conditions.

The dual-level CMI design addresses these conditions: category-conditioned constraints compress redundant variation within known classes; instance-conditioned constraints suppress augmentation-induced variation in unlabeled samples and, together with the base objective, retain the structure needed for category discovery. Conditional information thus provides a principle for organizing representation optimization under GCD's two supervision conditions.

These constraints share parameters with the base objective, and their gradient directions and scales can conflict. Nash-MTL coordinates them dynamically to address objective domination under fixed weighting, and preference qq explicitly controls the discrimination–discovery trade-off. The contribution links representation-imbalance diagnosis, dual conditional constraints, and dynamic coordination into a GCD refinement mechanism. The spectral ablation in Fig. 2, component ablation in Table 3, and multi-task weighting comparison examine these elements.

Q2. Default and discovery-oriented preferences

The default results in Tables 1–2 use q=1q=1, giving equal preferences to the base, category, and instance objectives. The category-level constraint directly promotes known-class compactness; the instance-level constraint supports novel-class discovery through augmentation invariance. On some stronger base methods, this symmetric preference produces a trade-off between higher Old and lower New accuracy.

Section 4.2 and Appendix G, Table 8 use the discovery-oriented preference q⋆=1.75q^\star=1.75, increasing the preference for the instance objective while task coefficients remain adaptive per batch. This preference is selected once using only Old-class validation data from RLCD/CIFAR-100, then fixed for RLCD, AptGCD, and all seven datasets. New accuracy improves over matched baselines in all 14 combinations, with statistically significant gains in all 13 held-out comparisons.

AuthorsOpenReview ↗

Follow-up on the Contribution and Preference Settings

Thank you again for your thoughtful comments. We hope our response clarifies InfoSculpt’s contribution and the distinction between the default and discovery-oriented preferences. Your feedback on these explanations would be much appreciated, and we remain available to discuss any outstanding questions.

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

InfoSculpt adds a 50-epoch refinement stage to pretrained GCD models. It combines class-covariance compression and augmentation-orbit losses with the host objective using Nash-style gradient balancing. A preference parameter shifts the balance toward novel classes. The experiments cover five hosts and seven datasets, including matched-seed comparisons for the new-oriented setting.

Critical strengths
  1. A refinement that improves several existing GCD methods without changing inference is potentially valuable. The host coverage and component ablations give this idea a reasonable empirical basis.
  2. Appendix G is a useful addition. It separates the calibration pair from the held-out comparisons and reports both novel-class improvements and old-class non-inferiority. In particular, reporting the latter as 11/13 avoids suggesting that every gain is free of an old-class trade-off.
Critical weaknesses
  1. Equations (8) and (9) describe different optimization problems. Writing A=GᵀG, the interior KKT condition for the stated simplex objective is Aᵀ(1/(Aα))=λ1. It does not give Aα=1/α. A simple counterexample is A=diag(1,4,9): the simplex objective has uniform optimal weights, while normalizing the fixed-point solution gives weights proportional to (1,1/2,1/3). Appendix D.6.3's normalization argument does not resolve this. The original Nash-MTL paper[1] derives the fixed point under a different constraint. This affects the claimed guarantees and leaves me unsure which objective was implemented.
  2. Several other details need correction before the proposed interpretation and gains can be assessed. Appendix D.3 drops the deterministic continuous encoder's conditional entropy as a constant without specifying a noise or discretization model that would make the Gaussian-KL bound finite. Section 4.2 also selects q*=1.75 from a listed grid that excludes 1.75. Finally, it is unclear whether the baseline receives the same extra 50 epochs. These are specific gaps in the derivation and comparison, rather than a general objection to covariance regularization.

[1] Navon, Aviv, et al. "Multi-task learning as a bargaining game." arXiv preprint arXiv:2202.01017 (2022).

Questions for authors
  1. Which objective and solver generated the reported results, including the asymmetric-preference runs? Would correcting the derivation change the updates?
  2. What noise or discretization assumption makes the CMI bound finite? If none is used, can the losses instead be justified as geometric surrogates?
  3. Please explain how q*=1.75 was selected and whether the matched baselines were continued for the same 50 epochs.
6 comments
AuthorsOpenReview ↗

We thank Reviewer RQJv for the constructive feedback.

Q1. Nash objective, solver, and correction

Thank you for identifying this issue. The error was an incorrect intermediate constraint. Training computes task coefficients using the fixed-point system in Eq. (9); the correction preserves this update rule and the corresponding experimental conclusions.

When deriving Eq. (8) from Eq. (7), we incorrectly constrained the coefficients to the probability simplex. The norm constraint should apply to the joint direction:

\qquad \|d\|_2\leq\epsilon,\quad g_i^\top d>0.$$ When a common descent direction exists, the KKT conditions, separating the overall update scale from the combination coefficients, give $$d\propto G\alpha, \qquad (G^\top G\alpha)_i=\frac{w_i}{\alpha_i}, \qquad \alpha_i>0.$$ The default symmetric preference $w=(1,1,1)$ recovers Eq. (9); asymmetric runs use $w=(1,1,q)$. Both settings use the stated iterative projection method with 20 numerical iterations per batch. We have corrected Eq. (8), the adjacent coefficient constraints, the theoretical statement in Appendix D.6.2, and the derivation in D.6.3. ## Q2. Finite geometric surrogates Our encoder is deterministic, with no additional representation noise or explicit discretization. Sections 3.2–3.3 and Appendices D.3–D.5 now identify the objectives as CMI-inspired geometric surrogates and explain their volume interpretation. Let $S$ be the empirical feature covariance and $E_S$ the ellipsoid with shape matrix $I+\rho S$. Then $$\frac12\log\det(I+\rho S) =\log\frac{\operatorname{Vol}(E_S)}{\operatorname{Vol}(B)},$$ where $B$ is the unit ball. Thus, the log-det loss measures the regularized volume of feature scatter. The identity matrix provides a fixed reference scale, and feature normalization bounds the empirical covariance, yielding a finite, computable objective. The category-level loss centers same-class features at their class mean and constrains their regularized spectral volume, forming compact known-class representations. The instance-level loss centers augmented views of the same image at their view mean and constrains augmentation-induced local scatter. Its squared-distance term measures average squared displacement from that mean, and its log-det term measures regularized volume. Together, they promote augmentation invariance and provide a geometric constraint for unlabeled data. ## Q3. Preference selection and matched training budgets Thank you for noting the omission. We have restored 1.75 to the candidate grid in Section 4.2. Appendix G specifies the selection rule: using only Old-class validation data from RLCD/CIFAR-100, select the largest candidate whose Old validation accuracy decreases by at most 0.5 percentage points relative to the default $q=1$. The selected $q^\star=1.75$ is fixed across both methods and all seven datasets. For RLCD and AptGCD in Table 8, the matched Base and InfoSculpt branches continue the same 200-epoch checkpoints for 50 epochs. Base retains the original objective throughout all 250 epochs; InfoSculpt adds its refinement objectives during the final 50. Total training steps, data splits, random seeds, and evaluation settings are matched. This protocol is specified in Appendices E.2 and G.

Thank you for correcting the Nash formulation and clarifying the calibration and matched training budgets. Two concerns remain. First, how does the solver handle batches without a common descent direction, and is convergence checked within the fixed 20 iterations? Second, the old-class validation constraint applies only to the calibration setting, while non-inferiority holds in 11 of 13 held-out comparisons. Please clarify the degradation in the remaining two cases and the limits of the claimed trade-off. Pending these clarifications, I maintain my rating of 2.

AuthorsOpenReview ↗

Clarifications on Nash Solver Implementation and Trade-off Limits

We thank the reviewer for the follow-up comments on the solver implementation and the scope of the trade-off. We address these points below.

Q1. Handling batches without a common descent direction and checking convergence

When a batch does not yield a feasible Nash solution, we follow the failure-handling principle of the original Nash-MTL implementation: retain the coefficients available before the solver failure and use them to combine the current batch’s gradients.

We will add a residual check based on Eq. (9), while keeping the number of numerical iterations fixed at 20 per batch:

rk=∥G⊤Gα(k)−w⊘α(k)∥2.r_k=\left\|G^\top G\alpha^{(k)}-w\oslash\alpha^{(k)}\right\|_2.

Here, w=(1,1,1)w=(1,1,1) for symmetric preferences and w=(1,1,q)w=(1,1,q) for asymmetric preferences. We will use a numerical tolerance of 10−310^{-3}, taking the setting in the original Nash-MTL implementation as a reference. If the residual does not fall below this tolerance after 20 iterations, we use the last available iterate.

Q2. The two comparisons that did not pass non-inferiority and the limits of the trade-off

The two comparisons that did not pass the old-class non-inferiority test are AptGCD on CIFAR-10 and ImageNet-100. Their mean old-class accuracies remained unchanged and increased by 0.1 percentage points, respectively; neither showed a decrease in the mean.

SettingOld: Base → InfoSculptDifferenceNew: Base → InfoSculptDifference
AptGCD / CIFAR-1094.4±0.5→94.4±0.294.4\pm0.5 \rightarrow 94.4\pm0.20.097.1→98.597.1 \rightarrow 98.5+1.4
AptGCD / ImageNet-10093.7±0.3→93.8±1.093.7\pm0.3 \rightarrow 93.8\pm1.0+0.182.2→90.982.2 \rightarrow 90.9+8.7

Accuracy is reported in %, and differences are in percentage points.

Both baselines already achieve high old-class accuracy, and the additional refinement produces limited old-class gains. In addition, the old-class preservation requirement originates from our preference calibration strategy. We calibrated the preference only on RLCD/CIFAR-100, then fixed q⋆=1.75q^\star=1.75 across the other methods and datasets. Different methods and datasets respond differently to the preference, so calibration on a single setting may not satisfy the same old-class preservation requirement in every setting.

This also defines the scope of our trade-off mechanism: InfoSculpt refines the feature space of an existing method, and its gains depend on the base model’s representations and the data distribution. A shared preference cannot guarantee old-class non-inferiority across all settings. To make the most of InfoSculpt, we recommend tuning the preference separately for each method–dataset combination.

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

This paper investigates the representation learning bottleneck in Generalized Category Discovery (GCD). Through singular value spectrum analysis, the authors observe that existing GCD models simultaneously suffer from over-compression and under-compression on novel classes, arguing that downstream fixes (e.g., debiased classifiers or pseudo-label refinement) fail to address this upstream issue. To tackle this, the paper proposes InfoSculpt, a post-training refinement framework derived from a balanced Information Bottleneck. It introduces a dual-level surrogate: a category-level loss on labeled data to compact within-class variance, and an instance-level multi-view loss on unlabeled data to preserve local structure while filtering augmentation noise. A game-theoretic weighting scheme is further employed to dynamically balance the gradients of the base loss and the two surrogate objectives. Experiments across five GCD baselines and seven benchmarks evaluate its effectiveness as a plug-and-play module.

Critical strengths

1.Analyzing the conflict between known and novel classes through the lens of the singular value spectrum and participation ratio provides a clean perspective, and the visualizations clearly illustrate the motivation.

2.Operating as a post-training refinement stage on top of converged checkpoints, InfoSculpt requires no architectural changes and introduces zero extra latency or memory overhead at test time.

3.Adaptive multi-objective balancing: Integrating a game-theoretic gradient balancing scheme automatically handles gradient scale discrepancies across multiple objectives, avoiding tedious manual hyperparameter tuning for loss weights.

Critical weaknesses

1.Disconnect between the core motivation and empirical results on novel classes: The paper criticizes existing GCD methods for merely patching downstream outputs and claims to resolve the root cause of novel-class failures by constraining upstream information allocation. However, the main benchmark results under default settings do not support this claim. Across modern baselines, the accuracy gains come almost entirely from boosting known (Old) classes, while novel (New) class accuracy remains stagnant or even degrades in nearly half of the settings (often further widening the performance gap between old and new classes). Even in the ablation studies, the full model yields lower new-class accuracy than the unmodified baseline. In practice, the method still heavily favors known classes rather than genuinely improving novel category discovery.

2.The actual technical formulation is overly simplistic and lacks solid grounding: Despite the heavy information-theoretic framing, the concrete technical realization is relatively weak. On unlabeled data, the instance-level objective merely enforces standard view invariance without cross-instance or category-aware guidance, and its multi-view volume penalty collapses into a simple monotonic re-weighting of pairwise distances under standard two-view settings, making the claim of pruning multi-dimensional non-semantic directions unconvincing. On labeled data, estimating per-class covariance matrices within standard mini-batches faces severe degeneracy on fine-grained datasets (e.g., CUB, Stanford Cars, Herbarium 19), where most classes have at most one labeled sample per batch.

3.Inconsistent evaluation and questionable baseline reproductions in the tuned preference experiments: To address the weak novel-class performance under the default setting, the authors introduce an asymmetric preference tuning parameter. However, the reproduced baselines for several frameworks under this setup are drastically lower than their officially published numbers, making the reported improvements less persuasive. Furthermore, relying on default settings for all main tables while falling back on tuned parameters to claim novel-class improvements makes the overall evaluation inconsistent.

Questions for authors

1.Since the paper is motivated by fixing representation failures on novel classes, how do the authors explain why the method mainly inflates old-class accuracy while leaving new-class accuracy largely unchanged or degraded across many modern baselines, frequently widening the old-new gap?

2.Given that only two augmentation views are typically used in practice, the proposed instance-level volume penalty mathematically collapses into a scalar function of the same pairwise distance. How can it independently act as an advanced mechanism to prune high-dimensional non-semantic directions?

3.With standard mini-batch sizes on fine-grained datasets with hundreds of categories, most known classes have zero or one sample per batch, making within-class covariance estimation degenerate. How is this term practically computed and stabilized in these scenarios?

4.Why do the reproduced baselines in the tuned-parameter experiments underperform their official published numbers by such large margins, and how do the default settings perform on those exact same reproduced checkpoints?

2 comments
AuthorsOpenReview ↗

We thank Reviewer BJPk for the constructive feedback.

Q1. Default Old/New gains

Thank you for raising this point. InfoSculpt aims to balance known-class discrimination and novel-class discovery in the representation space, helping existing GCD methods improve overall performance through latent-space regularization.

The default q=1q=1 assigns symmetric preferences to the base objective, category-level constraint, and instance-level constraint. The category-level constraint directly uses old-class labels to refine within-class structure, so gains on some base methods favor Old; decreases in New reflect the trade-off under this configuration. The parameter qq explicitly adjusts the preference for the instance-level task, allowing the same refinement framework to adapt this trade-off to discovery needs.

The preference experiments in Sec. 4.2 and Appendix G, Table 8 show that, with q⋆=1.75q^\star=1.75 favoring the instance-level task, New accuracy improves over the respective matched baselines in all 14 method–dataset pairs by 1.3–13.2 percentage points, averaging 6.07 points. Gains are statistically significant in all 13 held-out comparisons.

Q2. The instance-level term with two views

Thank you for pointing this out. With two views, the volume term indeed reduces to a logarithmic function of view distance. Its role is distance-dependent reweighting of the augmentation-consistency gradient. Let d=z1−z2d=z_1-z_2. Eq. (6) gives the two per-instance terms:

\qquad L_{\mathrm{inst}}^2=\frac12\log\left(1+\frac{\rho\|d\|^2}{4}\right).$$ Their gradients with respect to $d$ are $d/4$ and $$\frac{\rho}{4(1+\rho\|d\|^2/4)}\,d,$$ respectively. For a single view pair, both terms act along the same difference direction. Across a batch, the logarithmic term adjusts the relative weights of instances according to their view distances, changing the aggregated model update. Random augmentations produce different view differences across instances and iterations. Repeatedly constraining these variations suppresses augmentation-induced representation noise. The base GCD objective remains active and provides cross-instance structure learning and category discovery. The spectral ablation in the bottom row of Fig. 2 further shows reduced long-tail spectral noise in novel-class features after adding the volume term. Appendix D.5 provides the two-view reduction and gradient interpretation. ## Q3. Covariance with zero or one sample per class The category-level term uses labeled old-class samples in the current mini-batch and computes empirical covariance after centering within each class. Statistics are computed separately for each batch. Absent classes are skipped. For a class with one feature sample, both its centered covariance and its loss are zero; that sample still contributes to the base objective. Category-level training signals come from classes with multiple samples in the current batch. We use the identity-regularized covariance $I+\rho S_k$, which remains positive definite even when the empirical covariance is low rank, keeping the log-determinant finite and computation stable. Eq. (4) sums over classes present in the current batch; Appendix D.4 gives the definitions and boundary cases. ## Q4. Reproduction gaps and default results on the same checkpoints Thank you for pointing this out. The All and New columns for AptGCD and AptGCD + InfoSculpt in Tables 1–2 were swapped and have now been corrected. After correction, the published New accuracy on Stanford Cars is 53.6, versus our reproduced 51.8 in Table 8, a gap of 1.8 percentage points. For Herbarium19, the corresponding values are 42.3 and 40.7, a gap of 1.6 points. The reproduction gaps are approximately 1–2 points for AptGCD and 2 points for RLCD. For the default setting, both reproduced Base and InfoSculpt continue the same set of 200-epoch checkpoints for 50 epochs. Base retains its original objective throughout, totaling 250 epochs; InfoSculpt adds the refinement objectives during the final 50 epochs. Both use the same total number of training steps. Tables 1–2 report the default $q=1$ accuracies; Appendix G, Table 8 reports the matched Base and $q^\star=1.75$ accuracies. The default $q=1$ gains in New over the Base trained for the same number of steps are ($\Delta\mathrm{New}$, percentage points): | Method | CIFAR-10 | CIFAR-100 | ImageNet-100 | CUB | Cars | Aircraft | Herbarium19 | |---|---:|---:|---:|---:|---:|---:|---:| | RLCD | +3.1 | +2.3 | +3.3 | +2.2 | +3.2 | +1.7 | +4.4 | | AptGCD | +1.6 | +1.6 | +2.3 | +3.0 | +6.9 | +3.0 | +3.7 | With the same number of training steps, default $q=1$ improves New accuracy in 14/14 method–dataset pairs.
AuthorsOpenReview ↗

Follow-up on the Objective Formulation and Evaluation

Thank you again for your time and feedback. We would appreciate hearing whether our explanations of the Old/New trade-off, the two-view objective, and the covariance computation, together with the corrected and matched-budget results, address your concerns. Please let us know if any points require further clarification.

Reviewer_gKbWOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper looks into why current Generalized Category Discovery (GCD) methods struggle on novel classes. Through spectral analysis of novel-class features from a trained model, the authors show that two opposite pathologies, over-compression and under-compression, coexist in the same representation. They argue that the root cause is the unconstrained allocation between "discriminative" and "discoverable" information inside the representation, while existing remedies all act at the classifier output.

To fix this, the authors propose InfoSculpt, a post-training refinement module that further fine-tunes a converged base model. Starting from the Information Bottleneck, they derive a variational upper bound on conditional mutual information (CMI), which turns the allocation trade-off into an explicit, tunable knob, and under a Gaussian class-conditional assumption they obtain a tractable surrogate built around a log-determinant.

Critical strengths
  1. The paper is very clearly structured, moving from problem diagnosis to motivation to method design. Figure 1 performs SVD on the novel-class features of ProtoGCD on CIFAR-100 and gives a direct observation that two opposite pathologies coexist, attributing the old-new conflict in GCD to "unconstrained information allocation inside the representation.

  2. The method design is well grounded. Starting from the IB, the paper derives a variational upper bound on CMI, obtains a log-determinant surrogate under a Gaussian class-conditional assumption, and then instantiates it as a category-level and an instance-level loss. The two losses come from the same bound rather than being stitched together.

  3. Plugged into GCD, SimGCD, ProtoGCD, RLCD, and AptGCD, it brings improvements, without changing the backbone or the inference.

Critical weaknesses
  1. Weaker gains than concurrent plug-and-play methods, and no head-to-head comparison. As a plug-and-play method, InfoSculpt's improvements are smaller than those of recent plug-and-play methods from 2025 and 2026, namely Hyp-GCD (CVPR'25), APL (CVPR'25), and EAGC (CVPR'26), which focus on general representation learning, general local part learning, and gradient entanglement respectively, and report more promising results. Is there any head-to-head comparison with these methods, e.g., InfoSculpt+SimGCD versus Hyp-GCD+SimGCD, APL+SimGCD, and EAGC+SimGCD on the same base method? At present Hyp-GCD only appears as a standalone row in Table 2 and is not plugged into the same base methods for comparison.

  2. The choice of base methods is not justified. The main tables list more than ten methods, yet InfoSculpt is only added to five of them (GCD, SimGCD, ProtoGCD, RLCD, AptGCD). What is the rationale for choosing these methods rather than others, such as LegoGCD or FAIR?

  3. A key control is missing from the ablation. InfoSculpt trains a converged base model for another 50 epochs, with the base loss kept throughout. The Lbase-only row in the ablation table is identical digit-for-digit to the ProtoGCD numbers quoted in Table 2, which means there is no control where the base method is simply trained for 50 more epochs without the CMI terms. It is therefore unclear whether the gains come from the CMI losses or from the extra training itself.

  4. The wording is easy to misread. "Post-training refinement" and "zero overhead" can be read as training-free and cost-free, whereas the method is actually a 50-epoch full-model fine-tuning with 1.18× training cost

Questions for authors

Please to Critical Weaknesses, and Q1 is more important.

4 comments
AuthorsOpenReview ↗

We thank Reviewer gKbW for the constructive feedback.

Q1. Comparison on the same base method

We added comparisons using SimGCD as the common base. We also applied InfoSculpt on top of Hyp-GCD and EAGC to assess further representation refinement and report the results before and after integration.

InfoSculpt retains the base model's architecture and objective and refines existing checkpoints for 50 epochs. Adding InfoSculpt improves both All and New accuracy for Hyp-GCD and EAGC on all four datasets, supporting its complementarity with these methods.

APL was excluded because its code was unavailable.

All comparisons use equal training steps: both the base and InfoSculpt branches continue training for 50 epochs with the same number of optimizer updates. The three InfoSculpt configurations use q⋆=1.75q^\star=1.75. The current results are included in Appendix G, Table 8; we will complete the remaining entries as experiments finish. Entries below are All/Old/New accuracy (%).

ConfigurationComparisonCIFAR-100CUBImageNet-100Herbarium19
SimGCDBaseline (equal steps)80.10/81.20/77.8060.30/65.60/57.7083.00/93.10/77.9044.00/58.00/36.40
InfoSculpt + SimGCDOur reproduction81.16/82.25/78.9762.02/66.80/59.6384.64/93.66/80.1046.13/58.67/39.38
Δ\Delta (pp)+1.06/+1.05/+1.17+1.72/+1.20/+1.93+1.64/+0.56/+2.20+2.13/+0.67/+2.98
Hyp-GCD + SimGCDBaseline (equal steps)80.41/81.91/77.4159.01/61.31/57.8684.89/92.44/81.1044.63/57.07/37.93
InfoSculpt + Hyp-GCD + SimGCDOur reproduction81.35/82.69/78.6660.36/62.23/59.4386.98/92.34/84.2945.93/56.99/39.98
Δ\Delta (pp)+0.94/+0.78/+1.25+1.35/+0.92/+1.57+2.09/-0.10/+3.19+1.30/-0.08/+2.05
EAGC + SimGCDBaseline (equal steps)84.49/83.23/87.0266.95/67.65/66.6083.29/94.77/77.5247.98/61.43/40.74
InfoSculpt + EAGC + SimGCDOur reproduction85.65/84.24/88.4867.08/67.53/66.8583.90/94.64/78.5050.49/62.43/44.07
Δ\Delta (pp)+1.16/+1.01/+1.46+0.13/-0.12/+0.25+0.61/-0.13/+0.98+2.51/+1.00/+3.33

Q2. Choice of base methods

We selected the five base methods for their roles as established references and their diversity: GCD and SimGCD provide standard reference baselines; ProtoGCD and RLCD cover prototype and pseudo-label optimization frameworks; AptGCD covers prompt learning.

LegoGCD serves as a reference comparator in the main tables. For integration, we prioritized AptGCD, which uses LegoGCD as its classification baseline and adds prompt learning, to broaden mechanism coverage.

FAIR shares InfoSculpt's focus on representation refinement, so we prioritized the other base methods for integration.

Q3. Control for the additional 50 epochs

We updated the LbaseL_{\mathrm{base}}-only row with results from 50 additional epochs using the base objective alone. Both this control and InfoSculpt use a 50-epoch continuation budget.

On Stanford Cars and Herbarium19, InfoSculpt improves All by 2.28 and 2.39 pp and New by 0.59 and 1.96 pp, respectively, over Base-only continuation. This equal-budget comparison shows additional gains from the auxiliary objectives in these two settings. The updated LbaseL_{\mathrm{base}}-only row is marked in blue in Table 3.

MethodStanford Cars (All/Old/New, %)Herbarium19 (All/Old/New, %)
ProtoGCD (Base-only)53.14/72.61/43.7443.44/59.03/35.05
InfoSculpt + ProtoGCD55.42/78.39/44.3345.83/62.23/37.01
Δ\Delta+2.28/+5.78/+0.59+2.39/+3.20/+1.96

Q4. Terminology and computational cost

"Post-training refinement" is a 50-epoch training phase after the base model has converged. The reported approximately 1.18×1.18\times total computation is relative to a 250-epoch base-only run. "Zero overhead" refers to inference: the refined model retains the backbone, classification head, and forward path, with no additional inference computation or model storage. Training costs are reported in Appendix F.

AuthorsOpenReview ↗

Follow-up on Comparisons and Training-Budget Controls

Thank you again for your time and feedback. We would appreciate hearing whether our response, particularly the comparisons on a common base method and the equal-budget continuation controls, addresses your concerns. Please let us know if any points remain unclear; we would be happy to clarify them further.

Reviewer_gKbWOpenReview ↗

Response to the authors' rebuttal

The authors' response has addressed my concerns. In particular, the results for Q1 (Comparison on the same base method) are important. I strongly suggest placing them in the main text or adding them to Table 1 in the revision. This will fully demonstrate the effectiveness of InfoSculpt as a plug-and-play method.

AuthorsOpenReview ↗

We thank Reviewer gKbW for the positive feedback and helpful suggestion

We thank Reviewer gKbW for the positive feedback on our response and the helpful suggestion. Following your suggestion, we have incorporated the same-base comparisons into Tables 1–2 in the main text (p. 7). We have also completed the remaining experiments and made minor updates to the accompanying text. Thank you again for your constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.