acceptodds
Withdrawn from ICLR 2027

Energy Gradient Specification: Provable Generalization Guarantees for Learnware Identification and Reuse

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The learnware paradigm enables reusing well-established models on new tasks without retraining or exposing private training data, and a learnware dock system identifies reusable candidates through the compact specification stored with each model. However, existing specifications primarily capture statistical information, which may discard predictive information under tight storage budgets and make specification similarity an unreliable indicator of reuse compatibility. In an attempt to address this limitation, we propose Energy Gradient Specification (Egs), which characterizes and compares well-established models and user tasks in a shared one-step energy gradient space. Specifically, Egs optimizes compact proxy sets to match their gradient responses and uses the resulting gradients as specifications. We further establish a capacity-based generalization bound with only logarithmic dependency on the learnware size, together with an excess-risk decomposition explaining why reuse accuracy rises and then saturates as the dock system expands. Our theoretical results offer new insights into understanding how the learnware size affects reuse performance. Extensive experiments on image and text benchmarks show that Egs outperforms existing distribution-based specification methods across multiple learnware reuse scenarios.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Learnware Paradigm, Model Reuse, Generalization Analysis

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_p9vmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes EGS, a class-wise gradient-based specification for learnware identification and reuse. It uses gradient responses of a shared network to characterize learnware capabilities and matches specifications to select learnwares for new tasks. The paper also provides theoretical analyses, alongside experiments on image and text classification benchmarks.

Critical strengths
  1. The paper is clearly structured and provides explicit assumptions and detailed proofs for its theoretical results.
  2. The experiments cover image and text benchmarks, including reuse beyond original functionality, few-shot tasks, and corrupted learnwares.
Critical weaknesses

Weakness 1: The substantive advance of EGS over the existing PAVE specification and the motivation underlying its design have not been sufficiently established.

EGS and PAVE [1] are closely related in their use of signals induced by tasks or models in a shared model to characterize capabilities. EGS uses the class-wise one-step gradient responses of a shared network on the data used to construct the specification. PAVE uses the parameter changes of a public pre-trained model after multi-step adaptation on the corresponding data. Moreover, PAVE has been validated in capability characterization settings that include large language models, whereas EGS has primarily been evaluated on ResNet110 and TF-IDF-based classification models. From this perspective, under experiments with controlled models, tasks, and computational resources, the shared network and one-step gradient responses used by EGS are merely simplifications of the public pre-trained model and multi-step adaptation used by PAVE. However, in the limitations on page 31, the authors state, "Unlike PAVE (Shi et al., 2026), EGS does not rely on large-scale pre-trained foundation models or task-specific adaptation, making direct comparison under our controlled experimental protocol less informative." This does not establish a substantive contribution of EGS beyond the existing PAVE framework. The incremental contribution is limited to the class-wise construction of the specification, but similar ideas have already been adopted by prior works [2,3].

Meanwhile, the motivation that "under a limited specification size, distribution matching may discard class-conditional and task-relevant predictive information" does not adequately distinguish different issues. An exact KME with a characteristic kernel preserves information about the distribution it embeds. The absence of class information or information about model behavior from a marginal distribution representation is a different limitation from the compression error introduced by a finite reduced set. In Figure 1, class-wise distribution matching already substantially narrows the gap to gradient matching: under the smallest budget, CRKME and EGS retain 81.7% and 87.2% of the full-data linear probe performance, respectively. This result does not justify attributing all gains over RKME based on marginal distributions to the gradient mechanism, nor does it support a broad dismissal of distribution-matching specifications. Furthermore, "tight budgets" are primarily defined by the number of prototypes, whereas, with the probing network and number of classes fixed, the storage required by the final EGS gradient specification does not vary directly with the number of proxy prototypes. The paper does not establish a comparison under matched actual storage budgets and therefore does not sufficiently substantiate its motivation centered on tight storage budgets.

Weakness 2: The theoretical analysis lacks a substantive connection to the EGS specification, while missing key baselines and discrepancies between the paper and the supplied implementation leave its effectiveness claims insufficiently validated.

The theory establishes generalization guarantees for learnware identification and reuse, with a selection cost logarithmic in the number of candidates. It further explains performance saturation through dock coverage and specification distortion. However, the authors do not derive these key properties from the proposed EGS construction. Similar analyses would remain applicable to other specifications satisfying the same conditions. The theoretical contribution therefore lacks a substantive connection to the specification contribution: the theory neither explains why EGS is effective nor establishes its advantages over existing specifications.

The experiments do not sufficiently bridge this gap either. In addition to PAVE, Linear Proxy [4] and RNCME [5] are both directly relevant to the paper's emphasis on model functionality or predictive information. However, although the former is cited, it is not included as an experimental baseline. The latter is neither discussed nor compared against. Without these directly relevant competing methods, the current results are insufficient to establish the added value of EGS over existing specifications that account for model functionality.

The supplied implementation also shows apparent discrepancies with the specification construction and class matching described in the paper, as detailed in the Questions below. These discrepancies need to be clarified to establish whether the reported results correspond to the proposed EGS method and support the mechanism claimed by the authors.

References

[1] Hao-Yu Shi, Zhi-Hao Tan, Zi-Chen Zhao, Yang Yu, and Zhi-Hua Zhou. A Study on PAVE Specification for Learnware. ICLR, 2026.

[2] Peng Tan, Hai-Tian Liu, Zhi-Hao Tan, and Zhi-Hua Zhou. Handling Learnwares from Heterogeneous Feature Spaces with Explicit Label Exploitation. NeurIPS, 2024.

[3] Jian-Dong Liu, Zhi-Hao Tan, and Zhi-Hua Zhou. Identifying and Reusing Learnwares Across Different Label Spaces. IJCAI, 2025.

[4] Lan-Zhe Guo, Zhi Zhou, Yu-Feng Li, and Zhi-Hua Zhou. Identifying Useful Learnwares for Heterogeneous Label Spaces. ICML, 2023.

[5] Zi-Yu Mao and Ming Li. Identifying Learnwares via Reduced Neural Conditional Mean Embedding. ICML, 2026.

Questions for authors
  1. How are the submitted models' predictions incorporated into the implementation of Eq. (2)? The default objective="distribution" branch computes class-wise RFF mean embeddings, while the behavior branch uses fixed uniform targets. Which branch and configuration were used to produce the reported results?

  2. How is the partial one-to-one matching constraint in Eq. (5) enforced in the experiments? The provided matching routine independently takes class-wise minima and permits multiple user classes to select the same local class. Were the results obtained with this routine or with the constrained matching procedure described in the paper?

  3. The AI use statement says, "Generative AI tools were used solely to aid and polish the writing of this paper." Some portions of the submitted code give the impression of AI-assisted development, although this cannot be established from code style alone. Could the authors clarify whether generative AI was also used to assist with code development? (This question is for clarification only and is not a reason for my low rating.)

Reviewer_uA6HOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes learnware i.e reusing well established models on new tasks without re training. Current methods utilize statistical information which discards useful predictive information. A "dock system" stores many pretrained models, each with a compact specification, and a user with a small labeled dataset retrieves suitable models by comparing specifications. The authors propose Energy Gradient Specification (EGS). A shared, randomly initialized probe (a frozen feature extractor, a random Fourier feature map, and a random linear head) defines a common gradient space. For each learnware, class-wise gradients of the probe loss are computed on the developer's data, relabeled with the model's own predictions. A small synthetic proxy set is optimized to reproduce these gradients, and the gradients are stored as the specification.

Critical strengths

(i)The problem is practical: reusing trained models when training data is private and architectures are heterogeneous. (ii)A shared probe puts all learnwares and user tasks in one coordinate system, so specifications are directly comparable and cheap to compute.

Critical weaknesses

(i)The paper doest engage with prior works i.e The core mechanism, optimizing a small synthetic set so that its gradients match those of real data, is gradient-matching dataset distillation / coreset selection (e.g., Zhao & Bilen, 2021, which the paper cites). The paper does not explain how EGS differs from this line beyond the application. The differences that do exist (pseudo-labels from the model, a shared random probe, retrieval instead of student training) are never argued to matter, and related coreset and distillation work is cited sparingly and not engaged with.

(ii)The introduction motivates one-step gradients by citing Ba et al. (2022). As I understand it, that analysis concerns squared loss and first-layer feature learning in a high-dimensional proportional regime. EGS uses cross-entropy and differentiates only a random linear head over fixed features, so the cited result does not support the claim that these gradients carry task-relevant information. The paper also does not engage with related work on the low-dimensional structure of gradients (e.g., Gur-Ari et al., "Gradient Descent Happens in a Tiny Subspace"), which bears directly on what Eq. 2 captures.

(iii)The paper claims gradient matching preserves predictive information that distribution matching discards. Alternative explanations for the gains are never stated or tested: (a)class-conditional feature statistics already suffice (b)the gain comes from pseudo-labeling with the model's predictions rather than from gradients the random-feature kernel geometry drives retrieval (c)gradients concentrate in a low-dimensional subspace, so a tiny proxy set works for reasons unrelated to the paper's account the results are insensitive to the choice of loss There is also no comparison against a plain distillation or coreset specification, which is the most natural baseline. The empirical gain therefore cannot be attributed to the proposed idea.

Questions for authors

(i)How does EGS differ from applying standard gradient-matching distillation to each learnware's pseudo-labeled data and comparing the resulting specifications? Please include this baseline. (ii)Given that the probe is a linear head over fixed random features, how do the class-wise gradients differ from class-conditional kernel mean embeddings of the same features? What do gradients add? (iii)How much do results vary across random draws of the probe? (iv) Do conclusions also hold under other losses (e.g., squared loss, hinge)? How does this relate to the reliance on Ba et al.? (v) What is the effective rank of the stacked class-wise gradients, and how does it relate to the number of classes? (vi)What is the correlation between gradient error and downstream utility, to put next to Fig. 1(c)? (vii)Which parts of Theorems 1 and 2 would fail for a non-gradient specification? What is new relative to standard finite-class Rademacher arguments and existing coreset analyses? (viii)What privacy guarantee do the stored specifications give?

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

The paper addresses the learnware paradigm, where a dock of well-established models is searched through compact specifications so that a user can reuse a model without retraining or exposing private data. It argues that distribution-based specifications may discard predictive information, and proposes Energy Gradient Specification: each model is summarized by class-wise one-step gradients of a shared random-Fourier-feature probe on a distilled proxy set, with developer-side targets being the model's own predictions. Users compute the same object from labeled data, and a partial one-to-one assignment routes each user class to the nearest frozen model. The theory gives a capacity bound whose dock-size dependence is only logarithmic in the number of learnware candidates, and an excess-risk decomposition into a geometric coverage term and a specification-distortion term that explains a rise-then-saturate scaling curve. Experiments cover 12 benchmarks, 131 developer tasks, three reuse scenarios, and four specification baselines, with Friedman, Wilcoxon, and Holm testing; the method averages 0.710 against 0.634 for DALI. The supplementary code is a specification library without drivers, configs, or a README.

Critical strengths

The specification design is a genuine conceptual step for the paradigm. Computing the gradients against the model's own predictions makes the specification a sketch of capability rather than of training distribution, which is exactly the limitation the paper identifies in prior work. This is consequential because it is the idea that survives even if the current implementation does not, and because it is portable to other model-reuse settings.

The theory supplies the first account of dock scaling. Theorem 1 verifies line by line and gives a logarithmic dependence on the number of learnware candidates, and the coverage-versus-distortion decomposition predicts the saturation shape observed in the scaling study. This is consequential because the learnware program has needed a vocabulary for how dock size affects reuse, and the paper provides one even where its assumptions are strong.

Critical weaknesses

Theorem 2 does not cover the algorithm that ships. The statement bounds the predictor returned by Algorithm 3, but the proof analyzes the unconstrained per-class argmin selector. When two user classes share a nearest candidate, the constrained assignment displaces one class to a strictly worse model, so the relaxed selector's risk is a lower bound on the deployed selector's risk and the theorem bounds the wrong object. Remark 2 discloses the relaxation; the statement does not. This is consequential because the title promises provable guarantees, and the guarantee that attaches to the deployed method is the one that fails.

The load-bearing assumption is untested and no theoretical quantity is measured. Assumption 3, that gradient distance ranks per-class risk, is the method's central empirical premise and is testable in an afternoon on the 131-task market the authors built, yet it is never tested; the coverage term, the specification distortion, and the user-sample-size dependence are never estimated or fitted. The headline margin over DALI is mean-driven: four text datasets carry it, image datasets mostly tie, the median gain is 0.55, and the significance is borderline. This is consequential because the experiments validate the heuristic rather than the guarantees, and the empirical margin is a text-modality effect.

Questions for authors

Clarification requests (these could change my recommendation):

1- Can you restrict Theorem 2's statement to the unconstrained selector, or bound the displacement cost when the one-to-one constraint binds?

2- How often does the constraint bind on your markets, and what is the risk of the deployed selector relative to the relaxation?

3- What is the correlation between gradient distance and per-class risk for Assumption 3 across the 131-task market, and how often does the assumption fail?

4- Can you fit Figure 1(d) to the (1-p)^c coverage form and the log-c selection term, and report the parameters and residuals?

5- Can you measure the specification distortion by comparing proxy to full-data gradients at several proxy sizes?

6- Why are EGS, DALI, and Oracle identical in mean and standard deviation on four datasets in Table 1? Is the selection coinciding, and does the tie survive a second market draw?

7- How does EGS differ in mechanism and guarantees from probe-based transferability estimators (GBC, LogME, TransRate) and from the gradient-based PGE and ModelGiF?

8- Will you release drivers, configs, the matching solver, the evaluation harness, and the seeds? The current zip contains a library but no runnable pipeline.

Suggestions for future work (not needed for my assessment):

  • Add a train-your-own reference and a memory-matched specification baseline.

  • Report the runtime and construction cost of a specification against RKME and CRKME.

  • Add heterogeneous label-space baselines for the Beyond scenario.

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

Related papers

Discussion (0)

Sign in to comment.