Is the Shortcut in the Probe or in the Frozen Encoder?

A linear probe is the standard way to read out what a pretrained network "knows": you freeze the network and train only a single linear layer on its features. Probes are known to latch onto easy shortcut cues, such as a pasted-on digit or a background color, while ignoring harder, more reliable ones. The usual explanation blames the probe and its training. A separate line of work shows that network architectures prefer simple functions even before training: deeper networks squeeze their features into fewer effective dimensions, and Transformers favor outputs that change little when inputs are perturbed. This note proposes to test a specific connection between the two. A probe trained by logistic regression ends up preferring whichever cue the frozen features separate with the widest margin. If the encoder's built-in preference for simplicity gives simple cues large, high-variance directions and pushes complex cues into small ones, then most probe shortcut behavior is inherited from the encoder's geometry rather than created by the probe.

The claim is useful only if it predicts something. We propose a measurable quantity, the margin tilt: how much more widely the frozen features separate the simple cue than the complex cue, with each cue measured on its own. We will test whether this tilt, measured before any joint probe is trained, predicts how much the joint probe relies on the shortcut, across encoders that differ in depth, attention structure, initialization regime, and pretraining. The same account predicts which repair should work. If the complex cue is still linearly present but sits in low-variance directions, then reweighting directions should recover it; a regularized whitening of the features does exactly that. If the complex cue has been erased, no change to the probe can help. Sorting probing failures into "present but outweighed" and "erased" is the deliverable, and it decides when hardness-promoting probes are worth using at all.

Core Claim

Logistic regression on linearly separable features converges in direction to the maximum-margin separator. On a frozen representation $z = f_\theta(x)$, the probe's choice between two predictive cues is therefore set by the margins each cue achieves in the encoder's own geometry. Let $\gamma_s$ and $\gamma_c$ be the normalized maximum margins of probes trained on data where only the simple cue, or only the complex cue, carries the label, with the other cue randomized. The margin tilt is

$$\tau = \log(\gamma_s / \gamma_c).$$

We hypothesize three things. First, shortcut reliance in the joint probe increases monotonically with $\tau$ across encoders. Second, $\tau$ is driven largely by architectural simplicity bias, so it grows with depth and lower effective rank and appears in untrained networks in the ordered regime but not in the chaotic regime. Third, the family of whitened features

$$z_\alpha = (\Sigma + \lambda I)^{-\alpha/2} z,$$

where $\Sigma$ is the feature covariance on the probing set, $\lambda$ is a shrinkage constant, and $\alpha \in [0,1]$ moves from raw features to fully whitened ones, lowers $\tau$ and shortcut reliance together whenever $\gamma_c$ is clearly above zero. It should do nothing when $\gamma_c \approx 0$.

Several papers motivate this. The Pitfalls of Simplicity Bias in Neural Networks shows that trained networks become invariant to complex, fully predictive features: randomizing them had "zero effect" on logits, and models reached only ~95% on $\hat{LMS}$-7 even though a 100%-predictive feature was available. Their explanation is about training dynamics. We ask how much of the same effect a frozen encoder hands to a convex probe. The Low-Rank Simplicity Bias in Deep Networks shows that depth pushes features toward low effective rank, measured as the exponentiated entropy of normalized singular values. It never asks what a linear readout can still recover under that compression; we turn that rank story into a prediction about probe margins. Simplicity Bias of Transformers to Learn Low Sensitivity Functions reports CIFAR-10 sensitivity of 0.0050 for ViT-small against 0.3673 for DenseNet-121. This gives a second architectural axis, attention versus convolution, along which tilt might vary. Deep neural networks have an inbuilt Occam's razor supplies a method we reuse: adjusting the initialization weight variance $\sigma_w$ moves a network from an ordered, simplicity-biased regime to a chaotic one without changing its width. That lets us switch the architectural prior on and off at fixed dimension.

The closest prior work is A Closer Look at Model Adaptation using Feature Distortion and Simplicity Bias. It shows that linear probing followed by fine-tuning inherits the probe's shortcut, and it repairs this with hardness-promoting probes based on virtual adversarial training, uncertainty-driven perturbations, and averaged sparse probes. It treats simplicity bias as a property of probe optimization and reports that the repair helps much less in "low-distortion" tasks like Living17. Our proposal differs on the cause. We place the bias in the encoder's geometry, measure it before training the joint probe, and predict in advance which tasks a probe-side repair can help.

Research Program

The first component builds a set of encoders in which architectural simplicity bias varies in a controlled way. We take a pretrained ViT and ResNet. We add depth-overparameterized versions, made by inserting linear layers, which The Low-Rank Simplicity Bias in Deep Networks shows lower effective rank without adding capacity. We also include untrained versions of each architecture swept across $\sigma_w$ from the ordered to the chaotic regime. The untrained encoders matter for the logic of the study: they produce tilt with no pretraining at all, so any link between tilt and architecture cannot come from what the pretraining data contained. For every encoder and task we record effective rank, input sensitivity, $\gamma_s$, $\gamma_c$, and $\tau$.

The second component trains joint linear probes on cue-conflict data and measures reliance. Reliance is the accuracy drop when the simple cue is randomized, minus the drop when the complex cue is randomized. This follows the coordinate-randomization test of The Pitfalls of Simplicity Bias in Neural Networks. The central analysis is a rank correlation between $\tau$ and reliance across all encoders, with effective rank and sensitivity as candidate upstream causes.

The third component runs the whitening path. For each $\alpha$ we recompute $\tau$ and reliance. We then classify each encoder–task pair as "present but outweighed" if $\gamma_c$ is clearly positive and whitening lowers reliance, or "erased" if $\gamma_c$ is near zero and nothing at the probe level helps. This answers the open question the research file raises about depth: whether rank compression helps the probe by projecting away noise, or destroys complex features before the probe sees them. We will also check whether tilt measured on the probing set predicts Living17-style tasks where hardness-promoting probes gave only marginal gains.

Baselines

The naive reference point is a standard logistic-regression probe on raw frozen features. Its weight decay will be tuned on in-distribution validation data with the same search budget given to every other method. This is the probe whose shortcut reliance we are trying to explain, and every method shares its features, data splits, and training-set sizes.

The strongest existing method is the hardness-promoting probe family from A Closer Look at Model Adaptation using Feature Distortion and Simplicity Bias: LP(VAT), LP(UDP), and LP(Soup). We evaluate each both as a standalone probe and as initialization for fine-tuning. The reported number to beat comes from that paper: hp-LP+FT variants improved Randomized OOD accuracy on Dominoes by over 20% relative to standard fine-tuning. For their CIFAR-10 corruption and CIFAR-10.1 results, the file reports only a "noticeable boost" with no exact value. We will therefore rerun their methods ourselves on the same backbones and give them the same number of hyperparameter configurations as our whitening sweep over $\alpha$ and $\lambda$. Beating them has two parts. Whitened LP+FT must match or exceed hp-LP+FT on Dominoes Randomized OOD and on CIFAR-10-C mean corruption accuracy. Tilt must also predict, before training, which tasks show gains from either method, which hp-LP alone cannot do.

The key ablation separates geometry from regularization. It is a probe trained on randomly rotated features whose covariance spectrum is matched to the whitened features but whose principal directions are not aligned with the encoder's own. If this ablation repairs reliance as well as true whitening does, our story about the encoder's specific directions is wrong and the benefit is generic conditioning. A second control replaces each untrained encoder with a Gaussian random projection of the same output dimension and the same effective rank. This separates architectural structure from the effects of low rank alone.

Why not just…?

Why not just use a nonlinear probe? A small MLP might find the complex cue even when it is weak in linear geometry. But Probing as Quantifying Inductive Bias shows that expressive probes can learn the task themselves: under Bayesian model evidence, adding probe depth on random representations collapses evidence. A Non-Linear Structural Probe gains 8.44 attachment-score points in Tamil with an RBF kernel probe, but its kernels have structural limits of their own. We will still run an MLP probe. If it removes shortcut reliance even on pairs we classify as "erased," that would show the cue survives in a nonlinear form, and that our "erased" label means only "erased for linear readout." We will report it that way.

Why not just use hardness-promoting probes and skip the diagnosis? They are the main competitor and may match whitening on accuracy. What they lack is any way to tell in advance when they will help; their own Living17 result shows the gain varies by task. If tilt fails to predict where hp-LP helps, the diagnostic part of this proposal fails, and the experiments are designed to show that.

Why not just retrain the last layer on group-balanced data? If the shortcut is decorrelated from the label in the probing set, any probe will stop using it. This is a real threat, and it works where such labels exist. The point here is the setting without them: whitening needs only unlabeled feature covariance, and tilt needs only cue-randomized probes, which can be built from synthetic interventions. We will include balanced-data retraining as an oracle upper bound and report how much of that gap whitening closes.

Why not attribute everything to random high-dimensional projection? A reviewer could argue that untrained features are linearly separable only because random high-dimensional maps tend to be. The $\sigma_w$ sweep holds dimension fixed, and the rank-matched Gaussian projection control holds spectrum fixed. If tilt is the same in the chaotic and ordered regimes, or the same for the architecture and for the matched projection, the architectural part of our claim fails.

Proposed Experiments

The decisive synthetic test uses LMS-k and MNIST-CIFAR from The Pitfalls of Simplicity Bias in Neural Networks and Dominoes from A Closer Look at Model Adaptation using Feature Distortion and Simplicity Bias. In each, the simple and complex cues can be randomized independently, so $\gamma_s$, $\gamma_c$, and reliance are all exactly measurable. The prediction that would most clearly fail is this: across all encoders, including untrained ones in both regimes, the Spearman correlation between $\tau$ and reliance should be strongly positive, and $\tau$ should rise with depth and fall when moving into the chaotic regime.

The natural-shift test uses CIFAR-10 with CIFAR-10.1 and CIFAR-10-C, Living17, and DomainNet-Sketch. Here the cues are not labeled, so we use tilt on paired synthetic interventions, such as background or texture swaps on the probing images, as the predictor. Outcomes are OOD accuracy and mean corruption accuracy for raw LP, whitened LP, the hp-LP variants, and each of these followed by fine-tuning.

Three results would decide the question. The first is whether whitening lowers reliance on "present but outweighed" pairs and leaves "erased" pairs unchanged; if it helps both equally, the geometric account is wrong. The second is whether the spectrum-matched random-rotation ablation falls short of true whitening; if not, the effect is generic conditioning rather than the encoder's own directions. The third is whether low-effective-rank encoders cluster in the "erased" category under extreme depth, which would answer the depth question raised by The Low-Rank Simplicity Bias in Deep Networks.