acceptodds
Under review as a conference paper at ICLR 2027

The Reference Matters: Displacement Probing for Weight-Space Learning

Abstract

Weight-space learning predicts a model's attributes directly from its parameters, and probing-based methods do so efficiently by reading a model's responses to learned probes. Prior work shows that probing succeeds within a Model Tree—models fine-tuned from a shared pretrained origin—but not why. We trace this to the shared origin itself: since a tree's leaves start from the same weights, what distinguishes them is their displacement from that origin. This motivates probing the displacement rather than the raw weight, an intervention we call the DiSplacement Probe (DSProbe). We develop this idea along two axes: extending it to second-order Gram responses, whose displacement we show encodes the learned features induced by fine-tuning; and treating the reference as a design choice, learning a shared component to subtract beyond the pretrained origin. Across four architectures on the Model Jungle benchmark, the resulting Refined-DSProbe (R-DSProbe) sets a new state of the art, matches prior accuracy with substantially fewer probes, and supports a single encoder shared across Model Trees, reducing metanetwork parameters by 3.61 times.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
weight-space learning, neural network probing, scalable model analysis, model property prediction

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_CDp6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles the task of learning on neural network weights. While previous methods use predictors with only a multiplicative probing component or did not normalize the inputs, this paper shows that subtracting a constant matrix from the layer, prior to learning to probe it improves the accuracy. Most of the gain simply comes from normalizing the matrixor using the value of the pretrained weights, although learning it is better.

Critical strengths

The task of predicting the function of a neural network just given a single layer is interesting, and can make understanding large neural populations far more efficient. The paper presents a very simple yet successful approach. It is super easy to implement. It makes perfect sense to me that centering the data prior to fitting an affine predictor should work. While it may seem obvious, previous approaches did not do it, so it is novel in my view. The empirical results are good. This appears to work. It is compared on suitable benchmarks and against suitable benchmarks.

Critical weaknesses

I think this paper is justifying its approach in an incredibly complex way. First, it used a general formulation of high moments of the matrix when it is only using first a second moment. Second, it is spending about 2 pages describing the training dynamics of finetuning, justifying something that is fairly obvious i.e, that function vectors are an useful quantity, which is very well established. Furthermore, the best results are obtained by learning the offset, and this does not seem too related to the theory. Furthermore, centering in machine learning is very well established, that a simple reference would suffice. Overall, I think the writing is a blocker. I think we need a occam’s razor, simplest story as opposed to a complex story and proofs that are not necessitated by the evidence. Specifically, if the story really is to predict layer attributes we first center the first and second moments, and then learn an affine classifiers then so be it! I can imagine the authors feared this may sound too simple, but this reviewer things that given that previous approaches failed to do it, it is novel. Also, the SoTA performance suggests it is significant. To me it would be compelling and simple. It will also surely be used by future approaches, as it is so simple.

As this is ICLR, the revision can actually make these modifications and still be reviewed. I would be ok even if the change is rather large, although note the AC and other reviewers may take a different view. While my pre-rebuttal rating is weak reject, I like the result and give a good revision I could be persuaded.

Questions for authors

Can you ake the story sharper and simpler? s there justification for the high complexity?


The authro revision improved the clarity of the paper and I can raise my rating to weak accept.

1 comment

We thank the reviewer for reviewing our revision and for raising the score. Following the reviewer's suggestion, we simplified the story to match what the evidence supports, which substantially improved the clarity of the paper.

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

This paper tackles the discriminative tasks of weight space learning using probing based methods. Instead of directly processing all model weights, probing based methods pass data samples (probes) through the model and represent the model by its response. Prior work probed the raw weights of the models with learned probes instead of just data samples, single-layer probing, second order Gram matrices. The author proposes two new probing approaches, DSProbe & R-DSProbe, which instead probe the displacement of weights from a shared reference such as in Model trees. Arguing that the displacements remove common components of models in a model tree and thus increases the discriminativeness of the models.

This is evaluated on two tasks using Model Trees from the Model Jungle Benchmark. The first being predicting the subset of classes from CIFAR-100 a vision model was trained on, and the second is predicting which imagenet class a LoRA is trained for. The results showed a new SOTA for all tasks evaluated.

Critical strengths

The paper introduces a new probing approach that is relatively simple yet effective. Achieving SOTA across 4 different vision architectures on the Model Jungle Benchmark. It has a good motivation behind the methodology, which however could be augmented with some task-vector literature discussion.

Secondly the paper does provide extensive ablation studies regarding the choice of reference, higher order probing, layer selection etc. Which helps isolate the performance gain as the reference subtraction.

Critical weaknesses

Generalization. In this paper the experimental evaluation is restricted to the Model Jungle benchmark, which just regards Model Trees, meaning only models trained from shared pretrained checkpoints. Thus it remains a question in regards to how this can translate to more general model zoos. And even when restricting to Model Trees, the paper does not explore generalization to unseen Model Trees.

Methodology. Representing models through the displacement is established in task-vector literature (such as Editing Models with Task Arithmetic), but this connection is not discussed in the Related works. Secondly it seems that much of the performance gain comes from the subtraction of pretrained weights (DSProbe and the centering) and not much more by learning the reference (R-DSProbe).

Questions for authors
  1. Some of the improvements over MVProbe are relatively small, especially in the “generative” tasks of the Model Jungle Benchmark. Are these results statistically significant, e.g. with for example a t-test?

2- As it seems that R-DSProbe is only modestly improving the results over DSProbe in terms of accuracy, are there some other advantages it provides, e.g better cross tree generalization or similar? Table 4 does show cross-tree probing, however it does not present results for the DSProbe?

  1. In the Appendix B.3 it is mentioned that the zero-shot checkpoints are selected using their validation candidate sets. Are the classes used in validation and testing the same? Could the authors further explain this, and if this does/do not introduce leakage?
2 comments

Regarding the Model Tree setting not being limited to Model Jungle. The basis of the 70-80% number relies on a small subset of models on Hugging Face that do indeed report a base model. Across all of Hugging Face this number is much smaller. However I still acknowledge the real word application of model trees.

I appreciate the clarification regarding the contributions of the learned reference, explaining the practical difference where the root is not always known of a model tree, and new results showing the, albeit low, statistically significant improvements over MVProbe.

I will maintain my already postitive score.

We thank the reviewer for the follow-up and for maintaining the positive score. We agree that the 70–80% figure applies only to models that declare a base model and does not represent Hugging Face as a whole. We will state this scope explicitly in the revised paper and avoid generalizing beyond it.

Reviewer_7Dj7OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors describe a new probing mechanism for weight-space learning to learn to predict the model's attributes on the ModelJungle benchmark. This proposed probing mechanism focuses on differences in weights in a tree of models learned from a common ancestor. The authors show that this mechanism is state-of-the-art on the benchmark.

Critical strengths
  • The method is well-motivated, easy to understand and mathematically grounded.
  • The mathematical background and the derivations are well-structured, well laid out, and follow a logical thread.
  • Results show good improvement from the state of the art on an established benchmark.
Critical weaknesses

None

Questions for authors
  • How sensitive is R-DSProbe in practice to having access to the pretrained root or to an accurate estimate of the Model Tree structure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.