What Linear Probes Miss:
Multi-View Probing for Weight-Space Learning

What a single linear probe misses, multi-view probing finds.

ICML 2026

Ulsan National Institute of Science and Technology (UNIST)
Equal contribution    *Corresponding author
MVProbe architecture: four-branch probing

Open-source platforms host millions of fine-tuned models, often without metadata. Weight-space learning aims to identify a model's training data directly from its parameters — but a single linear probe is fundamentally limited: distinct weight matrices can collapse to the same response. MVProbe resolves this by combining first-order row/column projections with second-order Gram-based views, balanced by a theory-grounded per-sample standardization. The result: state-of-the-art identification across every architecture in the Model Jungle benchmark, and a leap from 36% → 98% on hard SD1k LoRA identification.

Abstract

The explosive growth of open-source model repositories has created a Model Jungle, where checkpoints are frequently shared without adequate documentation or metadata. While weight-space learning offers a pathway to identify and analyze these models directly from their parameters, processing full-scale weights is computationally prohibitive. Probing-based methods have emerged as a lightweight alternative, extracting permutation-equivariant representations via learnable probe vectors. However, existing probing methods are limited by a single-view design: they capture first-order structures but fail to encode the rich, higher-order correlation patterns inherent in row–column interactions. To bridge this gap, we introduce MVProbe, a multi-perspective probing framework that synthesizes first-order signals with interaction-aware (Gram-based) views. Our approach is theoretically grounded; we analyze the scaling laws of different probing orders to derive a principled standardization and fusion strategy that ensures balanced contributions from all branches. On the Model Jungle benchmark, MVProbe consistently outperforms the state-of-the-art ProbeX across diverse architectures, including ResNet, SupViT, MAE, and DINO.

Method

Existing single-direction probes summarize a weight matrix \(\mathbf{X}\in\mathbb{R}^{m\times n}\) by a single first-order projection \(\mathbf{X}\mathbf{u}\). Such projections are inherently ambiguous: distinct weight matrices can produce identical probe responses, collapsing them to the same representation. They also miss higher-order structure such as the row–column correlations encoded in the Gram matrices \(\mathbf{X}\mathbf{X}^{\top}\) and \(\mathbf{X}^{\top}\mathbf{X}\).

MVProbe addresses this with four complementary branches. First-order branches use learnable probe matrices \(\mathbf{U},\mathbf{V}\) and produce \(\mathbf{XU}\) and \(\mathbf{X}^{\top}\mathbf{V}\) — direct row/column projections. Second-order (kernel) branches use probe matrices \(\mathbf{W},\mathbf{Z}\) and apply the Gram operators to produce \(\mathbf{XX}^{\top}\mathbf{W}\) and \(\mathbf{X}^{\top}\mathbf{XZ}\), which encode pairwise similarity structure between output and input neurons, respectively. Each branch is followed by per-sample standardization, a branch-specific MLP, and feature concatenation; a shared encoder and classifier head complete the model.

We motivate this design from two perspectives. From kernel methods, Gram matrices encode pairwise similarity that first-order projections cannot capture. From landmark-based representations in manifold learning, the four probe matrices act as learned reference directions and kernel-space landmarks that together coordinate the geometry of the weight matrix.

Theoretical Justification

Conceptual illustration of Theorem 1. With the same probe matrix \(\mathbf{U}\), two distinct weight matrices \(\mathbf{X}_1\) and \(\mathbf{X}_2\) may collapse to the same first-order probe response \(\mathbf{XU}\), making them indistinguishable. Passing them through the Gram operator \(\mathbf{X}^{\top}\mathbf{X}\) once more (second-order probing) separates them again — recovering the information lost to a single linear projection.

Theorem 1 — Expressiveness of Second-Order Probes

Let \(\mathbf{U}\in\mathbb{R}^{n\times r}\) be a probe matrix with \(\mathrm{rank}(\mathbf{U})=r<n\), and define the first- and second-order features \(\Phi_1(\mathbf{X}) := \mathbf{X}\mathbf{U}\) and \(\Phi_2(\mathbf{X}) := (\mathbf{X}^{\top}\mathbf{X})\mathbf{U}\). Then there exist distinct \(\mathbf{X}_1\neq\mathbf{X}_2\) with \(\Phi_1(\mathbf{X}_1)=\Phi_1(\mathbf{X}_2)\) but \(\Phi_2(\mathbf{X}_1)\neq\Phi_2(\mathbf{X}_2)\). Consequently, when \(r<n\), adding second-order branches separates weight matrices that are indistinguishable to first-order probing alone.

Theorem 2 — Transpose-Complement Non-Redundancy

Let \(\mathbf{U}\in\mathbb{R}^{n\times r}\) and \(\mathbf{V}\in\mathbb{R}^{m\times r}\) be probe matrices with \(\mathrm{rank}(\mathbf{U})=r<n\). There exist distinct \(\mathbf{X}_1,\mathbf{X}_2\) such that \(\mathbf{X}_1\mathbf{U}=\mathbf{X}_2\mathbf{U}\) but \(\mathbf{X}_1^{\top}\mathbf{V}\neq\mathbf{X}_2^{\top}\mathbf{V}\). Hence the column-side branch \(\mathbf{X}^{\top}\mathbf{V}\) supplies information that is not recoverable from the row-side branch \(\mathbf{X}\mathbf{U}\) alone — both first-order branches are needed.

Theorem 3 — Scale Imbalance & Standardization

For \(\mathbf{X}\) with i.i.d. \(\mathcal{N}(0,\sigma^2)\) entries and unit-norm probe columns, the expected scale ratio between the second- and first-order responses is \(\mathbb{E}\!\left[\|\mathbf{S}^{(2)}\|_F^2\right]/ \mathbb{E}\!\left[\|\mathbf{S}^{(1)}\|_F^2\right] = \tfrac{n(n+m+1)}{m}\sigma^2 = \mathcal{O}(n\sigma^2)\), so naive concatenation is dominated by the second-order branch. Per-sample standardization yields \(\|\tilde{\mathbf{S}}\|_F^2 = mr\) regardless of order, equalizing branch contributions and motivating the standardization block in MVProbe.

Headline Results

MVProbe vs ProbeX on the Model Jungle benchmark.

Discriminative — fine-tune class identification

92.2%
ResNet
vs ProbeX 81.6%
92.3%
SupViT
vs ProbeX 88.1%
81.6%
MAE
vs ProbeX 77.1%
78.3%
DINO
vs ProbeX 72.5%

Generative — Stable Diffusion LoRA identification

97.9%
SD1k In-Distribution
vs ProbeX 35.8%  ·  ~2.7× higher
98.0%
SD1k Zero-Shot
vs ProbeX 52.4%  ·  robust to unseen classes

Discriminative Results

Accuracy (%), mean±std over seeds 1–5 on the Model Jungle benchmark.

Method ResNet SupViT MAE DINO
StatNN 55.2055.8054.8355.69
ProbeGen 78.2778.4870.6861.26
ProbeX 81.61±1.29 88.08±0.39 77.11±0.14 72.54±0.18
ProbeX (×4) 87.16±0.26 90.33±0.31 77.26±0.12 73.25±0.21
MVProbe (Ours) 92.24±0.25 92.33±0.37 81.62±0.15 78.29±0.31
Layer-wise robustness comparison

Layer-wise performance comparison. MVProbe (solid lines) vs. ProbeX (dashed lines) across all layers. Shaded bands indicate the performance volatility. MVProbe maintains higher accuracy across most layers and shows less sensitivity to layer selection.

Generative Results (Stable Diffusion LoRA)

In-distribution and zero-shot accuracy on SD_200 and SD_1k LoRA adapters (mean±std over seeds 1–5, layer 46).

Layer-wise performance on SD LoRA

Layer-wise performance on SD LoRA. MVProbe (solid lines) vs. ProbeX (dashed lines) across all layers on \(\text{SD}_{200}\) and \(\text{SD}_{1k}\) (In-Distribution and Zero-shot). Shaded bands indicate per-layer volatility; red arrows mark the largest gains.

SD_200 — 200 ImageNet classes

Method In-Distribution Acc Zero-shot Acc
ProbeX 98.48±0.48 94.01±0.77
ProbeX (×4) 97.72±0.50 93.53±1.99
MVProbe (Ours) 99.80±0.00 95.53±0.65

SD_1k — 1000 ImageNet classes

Method In-Distribution Acc Zero-shot Acc
ProbeX 35.75±2.44 52.42±2.48
ProbeX (×4) 32.46±3.08 51.14±3.88
MVProbe (Ours) 97.88±0.37 97.96±0.29

BibTeX

Preliminary citation. The PMLR proceedings entry has not been published yet. We will replace this block with the official BibTeX (with page numbers and the canonical PMLR booktitle) as soon as it becomes available.

% Preliminary citation — to be replaced with the official PMLR entry once the
% ICML 2026 proceedings are published.
@inproceedings{heo2026mvprobe,
  title     = {What Linear Probes Miss: Multi-View Probing for Weight-Space Learning},
  author    = {Heo, Eunwoo and Seo, Kyeongkook and Yoo, Jaejun},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026},
  series    = {Proceedings of Machine Learning Research},
  publisher = {PMLR}
}

Acknowledgements

We thank the authors of ProbeX (Horwitz et al., CVPR 2025) for releasing their codebase, on which this work builds. We also thank the maintainers of the Model Jungle benchmark and the Stable Diffusion LoRA datasets for making weight-space research broadly accessible.