acceptodds
Withdrawn from ICLR 2027

Gradient Parameter Specifications for Learnware Identification and Reuse

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The learnware paradigm enables users to solve new tasks by identifying and reusing learnwares from a learnware dock system, where each learnware pairs a well-established model with a specification of its reuse-relevant capabilities. However, existing specification methods primarily summarize data distributions or feature statistics, limiting their ability to distinguish tasks with similar inputs but different predictive requirements. To address this limitation, we propose Gradient Parameter Specification (Gps), a learnware specification that encodes task-dependent discriminative structure and model-dependent predictive behavior through class-wise one-step gradient directions. Specifically, Gps optimizes a small proxy sample set to construct learnware specifications from developer tasks and model capabilities, analogously constructs user specifications from labeled user tasks, and identifies reusable learnwares through partial class-wise gradient direction matching. Our representation-level analysis characterizes how gradient responses depend jointly on input representations and predictive targets, and explains how class-wise responses mitigate cross-class cancellation. Extensive experiments on twelve image and text benchmarks across three reuse scenarios show that Gps achieves the highest average performance among data-centric specification methods.

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

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

This paper proposes GPS, which is a method to identify class-model mapping for reusing pre-trained models. Specifically, it uses one-step gradient directions as learnware specifications to address the ambiguity of existing data-centric specifications. Extensive experiments are conducted on image and text classification benchmarks.

Critical strengths
  1. The authors provide theoretical analysis to demonstrate the effectiveness of their proposed method.
  2. The paper evaluates the proposed method on several image and text benchmarks and considers three reuse scenarios.
Critical weaknesses
  1. I have concerns about the necessity of the proposed method. In general, the capabilities of pre-trained models (e.g., the classification categories) are provided by the model provider. Therefore, simple semantic information, such as class labels or task descriptions, could serve as effective learnware specifications.

  2. The proposed method has several limitations: (i) it is only applicable to classification models. (ii) It is restricted to closed-set classification tasks. (iii) The proposed class-wise routing relies on the class identity of test samples. However, such class information is unavailable for unseen test samples.

Questions for authors
  1. Can the proposed method adapt to models beyond classification models?
  2. Can the proposed method adapt to tasks beyond closed-set classification?
  3. Do we need the class information of test samples during inference? How can we do inference without such class information?
Reviewer_HdzFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Gradient Parameter Specifications (GPS) for learnware identification and reuse. Under a shared probe with a fixed feature extractor, it optimizes proxy samples to match class-wise gradient directions of a prediction head. These directions are represented by orthogonal polar factors. Developer specifications use model predictions as targets, while user specifications use ground-truth labels. Partial class-wise matching identifies learnwares for complementary reuse. Experiments cover twelve image and text benchmarks across three reuse scenarios.

Critical strengths
  1. Class-wise matching supports complementary reuse. The method can select different learnwares for different user classes. I find the explicit class correspondences useful for combining their capabilities. The shared probe also allows specifications from heterogeneous submitted models to be compared.

  2. The lightweight probe achieves competitive reuse performance. GPS uses a frozen feature extractor and computes gradients with respect to a small prediction head. This avoids task-specific backbone fine-tuning. The reuse results across twelve image and text benchmarks and three reuse scenarios are promising. The analyses of model corruption, proxy sample size, and runtime are also useful.

Critical weaknesses
  1. Limited novelty relative to PAVE. I see substantial overlap with PAVE in both motivation and method. Both construct model specifications from model predictions and user specifications from ground-truth labels, then compare them in a shared parameter space. GPS Sec. 3.3 also closely parallels PAVE Sec. 3.1, and the corrupted-model experiment revisits the same model-quality motivation. GPS adds head-only probing, proxy optimization, and class-wise matching. I view these as incremental extensions of the existing method.

  2. A direct comparison with PAVE is needed. I am not convinced by the reason given in Sec. 5.4 for omitting PAVE. Both methods address the same identification problem, and GPS also uses pre-trained features. Their differences in adaptation cost can be measured directly. I would like to see a comparison of reuse performance and computational cost under the same experimental protocol. If a fair comparison clarifies GPS's incremental contribution, I would be willing to raise my rating from 2 to 3.

Questions for authors
  1. Component contributions: I would find three controlled ablations useful: aggregated versus class-wise gradients; cosine similarity of raw gradients versus their orthogonal polar factors; and direct task gradients versus proxy-induced gradients. How much does each component contribute?

  2. Experimental reuse protocol: I would like to understand how R\mathcal R in Eq. (31) of Appendix A.3 is implemented in each scenario. How are the selected learnwares' predictions mapped to the user label space and combined? Are the scores calibrated? Is the routing or fusion learned? Do all specification baselines use the same reuse protocol?

  3. Scalability to large label spaces: With KK classes on each side, matching requires K2K^2 class-pair comparisons per candidate learnware, followed by assignment. I am curious how this scales at large KK. Could the authors discuss the complexity and report matching time as KK increases?

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

The paper introduces a new Learnware paradigm methodology, including discriminating between gradient responses of a randomly initialized CNN to a multitude of image classification models for the task of image classification.

Critical strengths

works to study model reusability for future tasks

Critical weaknesses

The abstract is largely incomprehensible because of uncommon term usage in it. It expects the reader to be well known about the niche of the Learnware paradigm, which I had to also read before even reading the abstract - when this paradigm is not a widely-known area of research and most papers on it are done largely by a single team. In fact, this paradigm has significant overlap with distillation, model merging, and model reuse works, which are not discussed. Even after that, the abstract is quite vague and provides little to no information. For example, "However, existing specification methods primarily summarize data distributions or feature statistics, limiting their ability to distinguish tasks with similar inputs but different predictive requirements." doesn't really mean much. Neither do the terms "task-dependent discriminative structure" and "model-dependent predictive behavior" through class-wise one-step gradient directions. They mean absolutely nothing by themselves. The authors are advised to simplify and elucidate the problem statement as well as the solution in the abstract.

Other than fundamental mathematically based concepts, the paper should be self-supporting and self-explanatory. For example, a Learnware dock system is mentioned on line 4. There is no proper definition of what this dock system refers to. This sentence could be written in a much simpler and clarifying way instead of using niche terms which have no mathematical definition. "Given a user task, the dock system identifies suitable learnwares by matching the user requirement against stored learnware specifications and reuses either a single learnware or multiple complementary learnwares when the task extends beyond the original functionality of any individual learnware" - in the sentence, you should explain what user tasks are expected to be, or what the domain of user tasks is. What do you mean explicitly by a task system? What do you mean explicitly by learners? Are you just referring to pre-trained models, or are there any property requirements in terms of size, data, etc., that the model has been trained on? How do you define user requirements? Are there any constraints on user requirements? What do you mean by Learnware specifications? In my opinion, the paper needs a significant rewrite if it is meant to be human-readable. Lines 53 to 55 are not clear. It's unclear what distributional discrepancy you are talking about. Why is this so abstract? Why can't you name the exact distribution metric you're using. Gradient response is suddenly introduced in line 59 without any explanation. In line 60, you "adopt" a CNN as a proxy network for what purpose? This sentence is incomprehensible - "Unlike data-centric statistics, gradients jointly reflect input-dependent representations and model dependent predictive targets, providing a functional signal from a single backward pass through the classification head." The methodology is similarly unclear and cognitively heavy to read, given how much LLM is likely used in it. So this seems like it is specifically for an image classification task, although it is not mentioned anywhere in the methodology. The features or outputs are taken from different models and given to this single CNN, which is backpropagatable, which I'm not sure about if it violates the constraints of a Learnware system. Basically trying to figure out differences between gradients with respect to the same input from different models? A term "polar factor" is used with citation to a paper by Bo and Wang, which doesn't really have that term.
I don't think the paper is ready in terms of how it's presented its ideas and methodology itself. I would recommend the authors refer to works like LoraHub and EigenLorax to study reusability and compare with your method.

Questions for authors

please get the work that draft you write reviewed by a senior author, ideally not working directly in this field.

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

Related papers

Discussion (0)

Sign in to comment.