Scoring Natural-Language Shortcut Explanations by the Bits They Save

Medical image classifiers often succeed for the wrong reasons. They detect a chest drain instead of a collapsed lung, or a child's shoulder growth plate instead of pneumonia, and the failure only shows up after deployment. To find these shortcuts today, one of two things has to happen. Someone must already have labeled the suspected confounder, or someone must squint at saliency heatmaps that are known to be unreliable. We propose a third route. A language model writes candidate explanations in plain English, such as "the model's pneumothorax score depends on whether a chest tube is visible." We then score each sentence by how many bits it saves when transmitting the classifier's own predictions to a receiver who already knows the true diagnosis.

The scoring rule borrows from neural entropy coding, the practice of compressing data by pairing a learned probability model with an arithmetic coder, which assigns roughly $-\log_2 p$ bits to each symbol. Suppose knowing a hypothesis lets the receiver predict the classifier's outputs more cheaply than knowing the ground-truth label alone. Then the classifier's behavior carries information about that hypothesis beyond the pathology, which is the signature of a shortcut. The hypothesis sentence itself is also charged its length in bits under a language model, so long, overfitted stories have to earn their keep. The result is a ranked list of English sentences about what a model relies on. Each sentence carries a number in bits that can be checked, and none requires a pre-labeled confounder.

Core Claim

Define a shortcut hypothesis as a short text $h$ together with a per-image attribute $a_h(x)$. The attribute is produced by asking a medical vision-language model the yes/no question implied by $h$, for example "is a chest tube visible?". Let $f(x)$ be the audited classifier's discretized output and $y$ the ground-truth label. The score is a two-part description length:

$$ S(h) = L_{\text{LM}}(h) + L_{\text{preq}}\big(f(X) \mid Y, a_h(X)\big) - L_{\text{preq}}\big(f(X) \mid Y\big). $$

Here $L_{\text{LM}}(h) = -\log_2 p_{\text{LM}}(h)$ is the code length of the hypothesis text under a fixed language model. $L_{\text{preq}}$ is the prequential code length, meaning the total bits an arithmetic coder spends when a small predictor is trained online on images seen so far and codes each next image's classifier output. A hypothesis with $S(h) < 0$ pays for itself: stating it plus using it compresses the model's behavior more than the diagnosis alone does.

The claim is that ranking hypotheses by $S(h)$ does three things. It recovers known and planted medical imaging shortcuts without confounder labels. Its magnitude predicts how much robust accuracy the model loses when the shortcut is absent. And it gives explanations whose faithfulness is measured rather than asserted.

This combines the three directions without new machinery. Entropy coding supplies the currency. Language generation supplies the hypothesis space and the audience-facing output. Shortcut learning in chest X-rays and dermatology supplies the regime and the ground truth. Language Modeling Is Compression established the method we reuse: it showed that a language model paired with arithmetic coding gives real, decodable code lengths, and that model-size overhead must be counted. We reuse that accounting to charge explanations for their length. We do not compress images at all.

Relation to Prior Work

Two lines of work motivate the gap by requiring labeled confounders.

Benchmarking Dependence Measures to Prevent Shortcut Learning in Medical Imaging removes a known confounder, patient sex, from latent features using mutual-information estimation, distance correlation, or adversarial training. Detecting shortcut learning for fair medical AI using shortcut testing tests shortcut use by sweeping how strongly a labeled attribute is encoded and correlating that with unfairness. Its central lesson, "encoding does not equal usage," constrains our design. It is why we measure information in the classifier's outputs, conditioned on the label, rather than information in its features. Both methods need the attribute column in advance. We generate the attribute from text.

Three works motivate the distrust of pixel explanations as auditing tools. Benchmarking saliency methods for chest X-ray interpretation found Grad-CAM 24.0% worse than radiologists in mean overlap and 76.2% worse on lung lesions. Agreement Between Saliency Maps and Human-Labeled Regions of Interest found similar misalignment in dermatology. Hidden Stratification Causes Clinically Meaningful Failures in Machine Learning for Medical Imaging showed pneumothorax AUROC falling from 0.94 with chest drains to 0.77 without. That drop is the kind of failure a good hypothesis should flag.

Several works connect two of our three directions but none connects all three.

Neural Image Compression and Explanation links compression to saliency masks, but its "compression" is spatial downsampling of background pixels. It has no entropy coding, no text, and no shortcut analysis. Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance uses the entropy of attribution maps as a heuristic to pick shortcut-prone samples. That is dispersion of a heatmap, not description length, and it requires human masks on synthetic graphs. Unlearning Spurious Correlations in Chest X-ray Classification removes the growth-plate shortcut using good and bad Grad-CAM prototypes, but only after a human has already identified it. AnatomiX grounds a radiology language model in detected anatomy to resist a spatial flip shortcut. Its text describes the image, not the audited model.

Finally, iLLuMinaTE turns post-hoc attributions into theory-guided language model narratives. It reports 89.52% user preference and 94.68% agreement between a GPT-4o auditor and humans on fidelity. However, fidelity there is judged by another language model. Our contribution to that line is a fidelity score that is a code length and cannot be talked into agreement.

Research Program

The system has three parts.

The first is a hypothesis proposer. A language model is shown the task, a radiology or dermatology vocabulary, and textual summaries of images where the classifier disagrees most with the label. It emits short candidate sentences about devices, text markers, patient age cues, positioning, acquisition view, and similar factors. A fixed hand-written list of candidates serves as a control.

The second is an attribute realizer. A medical vision-language model answers the question implied by each hypothesis, image by image, to produce $a_h(x)$. Where ground-truth annotations exist, we measure the realizer's accuracy directly. The realizer can itself be wrong or shortcut-prone, and its noise shrinks bits saved toward zero. This makes the method conservative, but it could hide real shortcuts, and we report that sensitivity.

The third is the coder. A small predictor, such as logistic regression or a two-layer network over $(y, a_h)$, is trained online over a fixed image ordering. Its predictive distribution drives an arithmetic coder over the discretized classifier outputs. The actual bitstream length is reported, and it is averaged over several orderings to reduce variance.

The final report lists top hypotheses as sentences with their bit scores. It also includes a counterfactual sentence stating the classifier's accuracy on the subgroup where the hypothesis is absent.

The analysis also has to handle a real confound. Suppose an attribute saves bits given $y$. It may be a spurious artifact, or it may be an unlabeled clinical subtype, such as large versus subtle pneumothorax, that the label schema merges. We treat both as findings worth surfacing, since the second is exactly hidden stratification. We separate them using the intervention check described below, not the bit score alone.

Baselines

The naive reference point is to audit nothing and report aggregate test AUROC. A companion reference scores randomly drawn vocabulary hypotheses, which shows the null distribution of bit scores.

The first strong competitor is a saliency-to-language pipeline. Grad-CAM maps are overlaid on the images, and a multimodal language model is asked to describe what the highlighted regions show. This is what practitioners would build first. It gets the same images, the same vision-language model, and the same number of hypotheses per model.

The second is ShorT from Detecting shortcut learning for fair medical AI using shortcut testing, run with ground-truth confounder labels. It is an oracle with access to information our method lacks. Matching its detections without labels counts as success.

The key ablation replaces prequential coding with a held-out conditional-information estimate: fit on half the data and report test log-loss improvement from adding $a_h$. A second ablation drops the $L_{\text{LM}}(h)$ term. All variants share data splits, realizer calls, and a fixed tuning budget for the small predictor.

For downstream mitigation, the strongest reported numbers come from the CheXpert sex-confounded effusion setup. There, MINE with true sex labels reaches inverted-test AUROC 87.7 against a baseline of 45.8 (Benchmarking Dependence Measures). On SIIM-ACR, Reducing Reliance on Spurious Features in Medical Image Classification with Spatial Specificity reports robust AUROC 0.46 with binary labels and 0.82 with dense segmentation. Beating the saliency-to-language pipeline means higher recall of the true shortcut in the top three hypotheses at equal false-positive rate on clean models. Matching MINE means feeding our discovered attribute to MINE in place of the true label and landing within a few AUROC points of 87.7. No reported result was found on Lacuna for label-free, text-level shortcut discovery on any of these benchmarks.

Why not just…?

Why not just ask a multimodal model to explain the Grad-CAM map? The map it reads is already too coarse and misaligned to localize small findings (Saporta et al.). A fluent description of a wrong map is still wrong, and nothing in that pipeline measures whether the sentence matches the model's behavior. This is our main baseline, and the planted-shortcut recall experiment settles the comparison directly.

Why not just compute conditional mutual information with a held-out classifier and skip the coding language? This is a real threat. Prequential code length is essentially cumulative online log-loss, so for a single hypothesis the two estimates may agree. We expect the coding view to matter in two places. First, when hypotheses differ in granularity, the charge for text length and predictor complexity stops rich attributes from winning by overfitting. Second, when data per subgroup are small, prequential coding avoids a wasteful split. If the ablation ranks hypotheses identically across all benchmarks, we will say that the contribution is the proposer-realizer-scoring pipeline and not the coding formalism.

Why not just slice performance by available metadata? Slicing works for recorded fields such as sex, age, or site. It is blind to chest drains, pen markings, and burned-in text, which are the shortcuts documented in SIIM-ACR, ISIC, and the Covid-19 growth-plate case. Our metadata experiment runs both methods and checks that our method recovers the metadata findings too.

Proposed Experiments

We first plant shortcuts with known strength. We paste synthetic text markers onto CheXpert images with a correlation to effusion swept from 0 to 1, and we reproduce the sex-effusion confound from Benchmarking Dependence Measures. The decisive check is that $S(h)$ for the planted hypothesis becomes more negative as correlation rises, and tracks the measured drop in inverted-test AUROC. For models trained on decorrelated data, $S(h)$ should stay at or above zero.

We then turn to natural shortcuts: chest tubes on SIIM-ACR, growth plates on the Covid-19 CXR set used by eXBL, and surgical markings on ISIC. We ask whether these appear in the top three sentences unprompted. Where annotations exist, we also report realizer accuracy.

The intervention check separates spurious artifacts from hidden clinical subtypes. For each top hypothesis, we retrain with the discovered attribute used as the confounder label for MINE or gradient reversal. If robust AUROC improves, the hypothesis named a shortcut. If it does not, and the attribute correlates with severity, we reclassify it as hidden stratification.

Finally, two radiologists and two dermatologists blindly rate the top sentences from our method and from the saliency-to-language pipeline for correctness and usefulness. We report whether their ratings agree with the bit ranking.

The project fails if planted-shortcut recall does not beat the saliency pipeline. It also fails if bit scores do not predict robust-AUROC loss better than the held-out-information ablation, or if clean models produce many strongly negative scores.