MedQA-MM: Shortcuts Behind Medical Visual Reasoning

A benchmark score credits final answers, but not the route by which an item can be answered. In medical multimodal multiple-choice questions (MCQs), this distinction matters because a correct answer can be supported by the intended image finding or by benchmark-preserved cues in the wording of answers, non-visual clinical text, visible image text, artificial annotations, or device/context artifacts. We call the resulting score-level overinterpretation reasoning inflation. Here, a route is an observable input path that can support answer selection, not a claim about the model's hidden cognition. Across six medical multimodal MCQ datasets, we separate candidate cues from behavioral evidence through prompt- and image-side audits, modality ablations, and matched repairs that preserve the medical target and answer key. In a 13-configuration open-model panel, full-input accuracy is 62.63%, while text-only and options-only settings achieve 53.96% and 29.71%, respectively. Removing length-gap, absolute/conspicuous, and spatial/prepositional cues lowers accuracy by 6.58, 3.50, and 4.77 percentage points. We also construct MedQA-MM, a 1,000-item shortcut-mitigated subset, where text-only and options-only accuracy fall to 5.21% and 12.33%. This does not imply that models never use images; it shows that medical image-reasoning claims require route-level evidence.

ICON Decomposition: Auditing Deep Neural Networks with Multivariate Variance-based Concept-level Explanations

Deep neural networks often exploit spurious associations, a failure known as shortcut learning. Auditing for shortcuts requires testing many candidate concepts, such as acquisition settings or demographics. Current concept-based explainability methods test one concept at a time, asking whether it is decodable from a layer. These methods flag concepts that are merely correlated with the outcome or with each other, and their scores are not comparable across layers or concept types. We introduce ICON decomposition, which quantifies how much of a layer's variance each concept explains given all other concepts and the outcome, and how much none of them explains, yielding layer-comparable, calibrated scores that suppress false positives. On simulated data, ICON recovers concept importance more accurately than seven existing methods. On skin-cancer models with inserted artifacts, it detects induced shortcuts where baselines report false positives. On two brain-imaging models, ICON's sparse explanations are validated by retraining and out-of-distribution tests.

Controlling for Omitted Variable Bias in Deep Neural Networks

Control variables are widely used in statistical modelling to account for omitted variable bias of known confounders. However, they have largely been underexplored in deep learning. This is surprising, given that deep learning models encode image-inferable covariates, such as demographic variables, into their predictions when these covariates are correlated with the outcome---a form of omitted variable bias referred to as 'shortcut learning'. While many existing confound-control or fairness methods try to restrict the correlation of such covariates with model predictions, we show that this fails to correct for omitted variable bias. We therefore propose a control variable approach for deep learning models, based on generalised additive modelling of the effects of model inputs and covariates. As flexible additive models can suffer from concurvity, we introduce an estimation procedure that refits the final layer of a pre-trained network to include covariate effects, using cross-fitting with ridge penalisation. We show how these effects can be orthogonalised with respect to covariates to exclude their mediated effects and that model predictions can be marginalised over the covariate distribution to control for their effect. This yields unbiased, interpretable predictions and offers flexibility to model the desired effects depending on the scientific or fairness objective. We verify our approach using simulated images, and demonstrate consistent estimation of true effects. Existing methods either require more data or fail to recover the true effects. We apply our method to real neuroimaging data with experimentally induced confounding, where it recovers prediction performance to near the level of a model trained on unconfounded data. Code is available at https://github.com/mpff/cocodeel.

Shortcut Learning in a Public Grape Disease Dataset: Annotation Granularity as a Modulator, Not a Cause

Public datasets for agricultural disease detection are usually judged fit for use from reported metrics, which say nothing about whether the annotation scheme is internally consistent. On one public grape disease dataset (3288 images, 11995 boxes, 6 classes), varying model capacity, input resolution and detection paradigm yields a test-set mAP50 range comparable to seed-to-seed noise, with the bottleneck at small objects across all five architectures. The finding lies on the data side: one class is annotated at whole-leaf level (median box area 43.16% of the image) while the other five are annotated at lesion level. On 5156 cross-species images containing no grape, 65.7% of the false-positive boxes fall into that one class, an over-representation of 13.41x relative to its share of the training annotations. Counterfactual retraining establishes a causal effect of granularity on the magnitude of the shortcut: shrinking only that class's boxes cuts its cross-species false positives by 66%, and a placebo control confirms the effect is specific to the manipulated class. A manipulation in the opposite direction, with criteria registered in advance, returns a negative result: coarsening the finest class to whole-leaf level (0.57% to 40.37%), matched in box count and share of annotations and with higher in-distribution AP, still leaves its cross-species false positives at zero boxes, while the unmanipulated original class holds 50.0% of them. Annotation granularity is therefore a modulator of this shortcut, not its cause: it can amplify or attenuate a sink that already exists, but cannot create one, and what fixes the destination remains open. We also give a granularity screening statistic requiring neither images nor training, and show airborne lesion-level detection to be optically out of reach. The failure mode is invisible to in-distribution evaluation.

Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance

Graph Neural Networks (GNNs) can solve prediction tasks by unintentionally exploiting shortcuts---that is, edges, nodes, and features that correlate with but are not causal for the prediction---which compromise their reliability in out-of-distribution tasks. We introduce XIGL, an architecture-agnostic human-in-the-loop strategy for removing such shortcuts from GNNs. Our key insight is twofold. On the one hand, reliance on shortcuts can be detected by inspecting GNN explanations. On the other hand, once made aware of such shortcuts, sufficiently expert users can provide tailored corrective feedback, which helps deconfound the model. XIGL supports any query strategy; however, since corrective feedback can be expensive to acquire, we develop an active learning strategy for prioritizing explanations that are more likely to display shortcut behavior, lowering annotation and cognitive costs. We showcase the effectiveness of XIGL, including both existing and proposed explanation-based strategies, on several GNN architectures. Our implementation is available online.

Discovery and Spatial Characterisation of Multiple Shortcut Groups for Auditing Vision Model Bias

Deep learning models trained on datasets with spurious correlations can achieve high average accuracy whilst relying on shortcut features that do not generalise out of distribution. Whilst out-of-distribution testing highlights subgroup performance disparities arising from shortcut learning, it does not localise the regions within images that are associated with it. Existing research mostly uses attribution maps from interpretability methods to understand the spatial nature of spurious correlations. For example, conditional alignment methods separate task-relevant evidence from evidence tied to spurious correlations by comparing attribution maps from a task model, a sensitive attribute model, and a bias-reduced reference model. This yields shortcut-aligned and task-aligned contribution maps for each image. However, existing methods aggregate these maps across the dataset, potentially masking recurring spatial shortcut patterns that occur only in subsets of images. We address this limitation by grouping per-image shortcut and task contribution maps into recurring spatial patterns using K-means and non-negative matrix factorisation, and visualising the resulting shortcut groups through contribution maps and representative examples. Across CelebA, CheXpert, Waterbirds, Camelyon17, and ISIC2019, and across ResNet and ViT models, the discovered shortcut groups reveal both shared and distinct spatial patterns of shortcut and task contribution, with varying subgroup composition and error rates, enabling targeted inspection of image subsets with higher error rates. We perform input occlusion and internal test-time interventions to show that masking or suppressing task contribution regions substantially degrades the model classification performance and propose a combined shortcut suppression and task amplification feature intervention approach which generally reduces performance disparities.

Look What the Probes Dragged In! Real-World Chest X-ray Shortcuts in MedCLIP

Vision-language models, such as contrastive language-image pre-training (CLIP)-based approaches, have reached state-of-the-art (SOTA) results in medical artificial intelligence. However, recent work reveals that CLIP-based models remain vulnerable to shortcuts. We investigate how real-world shortcuts manifest across different layers of the medical CLIP-based model, MedCLIP, and its vision encoder, a frozen ResNet-50. We attach 17 linear classification probes to the intermediate layers of the ResNet-50 and train them on three different dataset configurations and targets: NIH-CXR14 (pneumothorax) and PadChest (cardiomegaly and pneumothorax). This setup allows us to observe model behaviour during evaluation using subgroup-based calibration and layer-wise confidence curves. We find that the final linear probes achieve a high AUROC but poor calibration in the models. The layer-wise confidence analyses suggest that shortcuts emerge at different depths. Patterns consistent with localised shortcuts, such as drains, appear at later layers, while patterns consistent with diffuse shortcuts, such as scanner-specific noise patterns, emerge earlier, aligning with previous work. Finally, we conduct a manual analysis of the images, which reveals data quality issues in both NIH-CXR14 and PadChest. Our findings underscore that even SOTA models remain vulnerable to shortcuts, and the need for high-quality and well-annotated datasets to draw solid conclusions. Code can be found on our GitHub: https://github.com/nikodice4/MedCLIP_shortcuts.

UniMod: Enhancing Multi-Modal Medical Diagnosis through Cross-Modality and Within-Modality Alignment

Multi-modal learning combining medical images and clinical text is promising for disease diagnosis. However, standard multi-modal training leads to shortcut learning: models exploit the easier modality (e.g., diagnostic cues in text) while neglecting harder-to-learn features (e.g., subtle visual patterns). We propose UniMod, a framework that mitigates shortcut learning by requiring each modality to predict the diagnosis on its own. It supervises image-only, text-only, and multi-modal classification simultaneously, so each modality must extract diagnostic features. We add cross-modality alignment for knowledge transfer and within-modality supervised contrastive alignment over same-diagnosis patients. On Harvard-Glaucoma, UniMod reaches 0.850 AUC, outperforming OGM-GE and Gradient Blending by 1.6-1.8%; on CheXpert Plus, it reaches 0.966 AUC, surpassing them by over 5%. UniMod also extends to 5-class multi-label diagnosis without architectural change, improving mean AUC by 0.097 over CGGM.

PragMatch: Separating Pragmatic Incongruity from Cross-Modal Mismatch in Large Vision-Language Models

Large Vision-Language Models (LVLMs) have demonstrated strong performance on multimodal benchmarks, yet it remains unclear whether they genuinely reason about relationships between images and text or rely on superficial correlations, known as shortcut learning. This question is particularly important for multimodal sarcasm detection, where successful prediction depends on recognizing pragmatic incongruity rather than treating sarcasm as simple image-text mismatch. We introduce PragMatch, a controlled benchmark of 3,000 image-text pairs derived from MMSD2.0, including original sarcastic examples and constructed literal and hard-negative pairs. We identify influential shortcut cues through systematic masking and evaluate their impact through targeted injection experiments. Our results show that LVLM predictions are sensitive to lexical, OCR-derived and stylistic cues, with injected surface signals causing substantial changes in model predictions despite unchanged underlying image-text relationships. Our findings reveal limitations in current LVLMs while PragMatch provides a systematic testbed for evaluating multimodal pragmatic reasoning beyond surface-level image-text alignment.

Invisible Shortcuts: Why Vision Encoders Know Your Camera

Deep vision models exploit shortcuts, relying on cues that correlate with supervision signals. Prior work has focused on visible biases, such as object-background or texture correlations. We identify a different source of shortcut learning: invisible metadata traces embedded at the pixel level, for metadata such as image processing and photo acquisition. We hypothesize that large-scale semantic supervision, whether through categorical labels (ImageNet) or billion-scale captions (LAION), naturally induces metadata-semantics correlations during pretraining, leading models to convert low-level signals into predictive features. By introducing controlled metadata-semantics correlations, we show that stronger ones produce systematically higher sensitivity to metadata traces and larger performance degradation under metadata distribution shifts. We further explore mitigation strategies applied during and after pretraining that reduce sensitivity not only to targeted metadata but also to unseen ones, without sacrificing performance on downstream tasks. Metadata sensitivity also has a positive side: it partly explains the strong generated-image detection ability of some encoders, while its mitigation can improve out-of-distribution generalization. Code: https://github.com/ryan-caesar-ramos/visual-encoder-traces

Agents Catching Agents: Shortcut Cascades and Benchmark Gaming in Clinical Multi-Agent Systems

Clinical decision support is moving toward committees of language-model agents deliberating on a shared workspace. We ask whether such committees can be gamed by shortcuts, cues a benchmark rewards but a clinician would ignore. Across seven cohorts on six public datasets spanning text (MedQA-USMLE, MedMCQA, MIMIC-CXR reports), imaging (NIH ChestX-ray14, MIMIC-CXR-JPG, CheXpert) and tabular ICU records (SUPPORT2), Gemini committees resist these cues in isolation (flip 5-16%), yet a socially plausible shortcut spreads: when two peers assert the same wrong answer, the holdout under test adopts it in 38% of cases, as does a false "pre-screen" system flag, on both capability tiers. Of three oversight agents, a gate cannot separate adoption from honest agreement (false-positive rate 100%); a same-lineage judge reading only the transcript flags adoption on text (precision 100%, recall 93%) but collapses onto the gate in imaging; a referee that privately re-queries the holdout transfers to imaging (77-88% precision, 13-21% false-positive rate). Tripling a cue's visual salience does not move contagion, whereas a second peer voice raises it by half again. Gaming a hidden rubric is near-silent: only 1/10 text and 1/134 imaging drifters name the rubric they moved toward. What games a committee is social plausibility, and only a referee independent of self-report catches it. Code: https://github.com/criticaldata/benchmaxxing

When Oracle Conditioning Misleads Deployment: Conditioning-Availability Bias in Echocardiographic Segmentation

Conditional segmentation models may be trained and evaluated with auxiliary signals cleaner than those available at deployment. We study this protocol-level manifestation of shortcut learning and auxiliary-variable shift in phase-conditioned echocardiographic segmentation. The complementary gap pair measures loss on the deployable oracle-estimated pathway and probes sensitivity on the oracle-random pathway. On held-out CAMUS data, one strong-cyclic, oracle-selected run fails severely with estimated phase, while sensitivity to incorrect phase persists across three runs. On EchoNet-Dynamic, the current estimator remains usable, but random-phase testing reveals strong latent sensitivity. Deployment-aware checkpoint selection and phase perturbation reduce both gaps with little change in mean Dice. Exploratory subgroup analyses quantify variation across measured strata, and a downstream ejection fraction (EF) audit shows that recovering segmentation does not necessarily recover EF error or signed bias. Together, the gaps test whether oracle-conditioned performance survives the inference pathway actually available at deployment.

When Refusal Looks Safe: The Refusal-Cue Shortcut in Safety Guard Models

Safety guards are widely used to filter harmful content and are typically trained via supervised fine-tuning on labeled prompt-response pairs. We audit two widely used safety-guard training datasets, WildGuardMix and GR-Train, and find that among responses to harmful prompts, refusal expressions co-occur almost exclusively with unharmful labels. This imbalance motivates what we term the refusal-cue shortcut: inserting a refusal cue into a harmful response could flip the guard's verdict from harmful to unharmful. The shortcut affects not only guards trained on these datasets but also officially released models such as LlamaGuard3 and Qwen3Guard whose training data is undisclosed. It persists across response positions and is generally stronger in smaller variants within a family. To mitigate it, we adapt sparse complementary masking as a lightweight post-hoc intervention that identifies and suppresses a small set of shortcut-associated attention heads and MLP neurons without retraining. On two primary benchmarks, the intervention achieves an approximately 79% relative reduction in response-initial detection failures induced by refusal cues, while preserving standard detection performance. Although optimized using cues at a single response position, the suppression effect transfers to unseen positions and datasets, suggesting that shortcut manifestations across positions are partly mediated by shared internal components. Further analysis provides evidence that shortcut reliance and legitimate refusal recognition are partially functionally separable, as suppressing the shortcut broadly preserves the guard's ability to recognize genuine refusals.

Right Answer, Wrong Method: Shortcut Hacking Misleads the Evaluation of LLM Reasoning on Frontier Science Benchmarks

Scientific reasoning benchmarks typically evaluate large language models (LLMs) using final-answer accuracy. However, a correct answer does not necessarily demonstrate the reasoning capability targeted by the problem. We identify Solution Hacking, a failure mode in which an LLM reaches the correct answer through invalid shortcuts, such as numerical search, enumeration, guessing, or answer-first verification, without providing a valid task-targeted derivation. We systematically analyze this phenomenon across difficulty levels, scientific domains, and frontier models. Solution hacking increases sharply with benchmark difficulty, from 2.2\% on common problems to 28.3\% on Olympiad-level problems and 37.4\% on HLE. Moreover, 8.2\%-44.1\% of answers credited as correct across frontier models are identified as hacked solutions. We further develop expert-inspired anti-hacking strategies, including an automatic judge and a test-time instruction. The results show that suppressing shortcut behavior substantially reduces reported accuracy while having a smaller effect on correct and non-hacked accuracy. These findings reveal that answer-only evaluation can overestimate the scientific reasoning capabilities of frontier LLMs.

Analysis of the Shortcut Learning and Clever Hans Effect in CNN based ECG Image Classification

Deep learning models for ECG image classification may achieve high accuracy by exploiting non-physiological visual cues instead of ECG waveform morphology. Given the black-box nature of deep learning models, their promise of high predictive performance often remains insufficiently translated into clinical or real-world trust, interpretability, and actionable decision-making. In this study, we examine shortcut learning and Clever Hans effect in a publicly available ECG image dataset using convolutional neural networks. In process we have created six image-derived feature sets (FSs), FS1: raw full ECG images, FS2: cropped waveform-only images, FS3: waveform-masked metadata images, FS4: red-arrow artifact images for the myocardial infarction class, FS5: contrast-enhanced images for the abnormal heartbeat class and FS6: Gaussian-blurred images for the normal class. These controlled representations were used to test whether classification performance persists when waveform information is removed or when artificial class-specific artifacts are introduced. Shortcut retention score, prediction consistency and confidence divergence across Feature-Set Representations have been calculated to assess the transparency about the learning pattern. Along with factual results, average Integrated Gradients and occlusion sensitivity test results are presented to inspect whether model attribution focused on ECG-relevant waveform regions or on non-clinical artifacts. Performance changes across feature sets and attribution patterns were used to identify potential Clever Hans behavior. This study evaluates whether ECG image classifiers learn clinically meaningful morphology or shortcut cues introduced by report layout, metadata, contrast, blur, or artificial markers.

Style over Substance: A Shortcut Audit of Emotion-Description Preference Evaluation

Preference over model-generated emotion descriptions is emerging as a standard evaluation metric for multimodal emotion understanding, exemplified by the MER2026 MER-Prefer track on EmoPrefer. Such benchmarks assume that predicting the preferred description requires grounded cross-modal understanding of the video. We conduct a systematic shortcut audit of EmoPrefer using content-blind probes. A simple logistic regression using only description length and generator identity, without processing the text, video, or audio, performs comparably to LoRA-finetuned 7B text and audio-visual judges (65.8 versus 66.8 WAF on EmoPrefer-V2). Generator identity is recoverable from description text with 99.5 percent accuracy, every candidate pair contrasts two distinct generators, and the human preference labels agree with a fold-exclusive per-generator win-rate prior on 66 percent of the evaluated pairs. When the human label conflicts with this prior, trained judges still follow the style prior on 63 to 80 percent of the pairs. On a length-matched subset that neutralizes verbosity bias, the tested media configurations yield no statistically significant improvement, while an ODIN-inspired diagnostic that decouples the style shortcut leaves its content head near chance. These results do not imply that human preferences are inherently stylistic or that the descriptions contain no emotional information. Instead, they show that the current scores can be reached without verifying either description against the video. We recommend source-balanced pairing, strict length control, counter-stereotypical sliced reporting, and multi-annotator consensus for future cross-generator evaluations. Code is available at https://github.com/jiabingyang01/EmoPrefer-Audit.

Controlling Implicit Shortcut Reliance in L2 Spoken English Auto-markers

Increasingly, speech and language processing tasks take either audio or text directly rather than extracting features from these as the input to the classifier or regressor. Often these systems make use of complex, for example transformer-based, processes that have the ability to derive highly non-linear mappings between the input and the output. Unfortunately these systems can also learn ''shortcuts'' where the classifier is overly reliant on particular aspects of the input to yield the output. For the task of language proficiency assessment, this over-reliance can enable learners to increase their score by exploiting the shortcut rather than improving their ability. This paper introduces a novel training criterion that is able to reduce the classifier's reliance on shortcuts, thus for example limiting this option for malpractice in language assessment. This process is illustrated on two forms of assessment system, one based on the audio the other on the speech recognition text. The results show that, for both systems, there is higher correlations with features that could be exploited for malpractice than expected from the human reference, indicating an over-reliance on these features. By introducing the modified training criterion, this correlation can be reduced to be closer to the reference correlation.

Auditing Protocol-Level Shortcuts in Large Audio Language Model Judges for Speech Evaluation

Large audio-language models (LALMs) are increasingly used as automatic judges for speech evaluation. However, high agreement with human ratings does not guarantee that their verdicts are grounded in the audio. A judge may instead rely on specialist labels or reference data supplied by the evaluation protocol itself, taking a shortcut in place of listening to the audio. In this paper, we audit such protocol-level ``shortcuts'' in LALM judges across three common deployment protocols: feature-blueprint judging, where the audio is replaced by a structured text description of acoustic features, reference-conditioned judging, and pairwise A/B comparison. Across six judges and four attributes, we find that several LALMs rely on protocol-level shortcuts. For example, in feature-blueprint judging, incorrect specialist labels reduce five judges' emotion accuracy to 0.10 or below, and in concatenated A/B comparisons, Qwen3-Omni-Thinking often picks the same slot regardless of order swaps. These results indicate that aggregate agreement can overstate the validity of LALM judges unless the model and the evaluation protocol are assessed jointly, and that each model-protocol pair should be evaluated with a matched shortcut probe.

ReMoDEx: A Local-to-Global Relevance-Based Model Decision Explainability Framework for large-Scale Image Datasets

Deep learning image classifiers achieve strong predictive performance yet remain opaque in how decisions are formed. A model may predict correctly while relying on irrelevant cues, shortcut associations, peripheral structures, or device level artifacts instead of task relevant regions. On large scale datasets this opacity is especially problematic, since inspecting heatmaps one sample at a time cannot scale to thousands of predictions. We propose Relevance Based Model Decision Explainability (ReMoDEx), a framework for systematic, dataset scale assessment of model decision behaviour in image classification. ReMoDEx defines a stepwise pipeline: model inference, target class selection, relevance map generation, heatmap standardisation, similarity based grouping of patterns, cluster level interpretation, and spatial relevance assessment. Local methods GradCAM++, Integrated Gradients, Occlusion Sensitivity, and Layerwise Relevance Propagation are each combined independently with a single global module that summarises an entire set of relevance maps into a few decision strategy clusters, replacing sample by sample inspection with an automatic, scalable summary. To demonstrate ReMoDEx, we applied it to a VGG16 based classifier distinguishing COVID-19, Normal, Lung Opacity, and Viral Pneumonia. The classifier showed stable performance (86.27% test accuracy, 0.9624 test AUC). However, each explainer combined with the global module consistently produced two recurring strategies: central thoracic region decisions and border/corner sensitive decisions, indicating possible shortcut learning that conventional metrics could not reveal. Masked image validation confirmed that model confidence and predicted class changed when central or peripheral regions were occluded. ReMoDEx thus provides a scalable relevance based decision assessment framework and an essential complement to accuracy based evaluation.

When Benchmarks Mislead: Shortcut Learning, Length Confounds, and the Limits of Cross-Dataset Generalization in Multilingual Fake News and Sarcasm Detection

Cross-dataset generalisation is a fundamental requirement for deploying text classifiers in real-world settings, yet systematic evaluation across corpora from different sources remains uncommon in fake news detection and virtually absent in sarcasm detection research. This paper presents a unified empirical study of zero-shot cross-dataset transfer in three domains: Urdu fake news detection (FND), English FND, and sarcasm detection. For each domain, we fine-tune xlm-roberta-base on one corpus and evaluate it on a second corpus from a different source, comparing against TF-IDF baselines with Logistic Regression (LR) and Support Vector Machines (SVM). In Urdu FND (Ax-to-Grind vs. Notri-Fact), we identify a severe length confound in the Ax-to-Grind dataset, fake articles average 3.4 times more words than real articles, causing catastrophic A to B transfer collapse (macro F1 = 0.005) while B to A achieves F1 = 0.771. Extension to English FND (WELFake vs. ISOT) and sarcasm detection (TweetEval Irony vs. Sarcasm Corpus V2) reveals that such failure modes extend beyond Urdu, confirming that shortcut learning from distributional artefacts is a cross-lingual, cross-domain challenge in binary text classification. We provide a reusable diagnostic methodology, combining class-conditional length analysis, bidirectional transfer asymmetry, and predicted label collapse inspection, applicable across any binary text classification setting.

Association Restoration Test: Revealing Restorable Shortcuts after Unlearning

Association unlearning aims to disable learned label-attribute shortcuts while preserving task performance. Existing evaluations mainly measure output-level robustness or probe whether shortcut attributes remain readable in frozen features, but neither test determines whether a retained association remains functionally usable by the original classifier. We propose the Association Restoration Test (ART), a post-hoc diagnostic for functional shortcut restorability. ART estimates class-conditional association directions, amplifies residual components, and evaluates the modified features with the original classifier head. Across Waterbirds, CelebA, SpuCoDogs, and an ISIC timestamp-artifact extension, we show that output metrics, representation probes, and ART characterize distinct aspects of shortcut mitigation. These findings motivate restoration-aware evaluation for unlearning and shortcut-mitigation methods that target learned associations rather than individual classes or concepts.

Shortcut Learning in Legal Judgment Prediction: Empirical Evidence from the UK Employment Tribunal

Current Legal Judgment Prediction (LJP) is constrained by its reliance on post-hoc judicial materials, increasing the likelihood that models perform retrospective classification rather than true forecasting. This paper empirically investigates shortcut learning in this context by studying claim-level outcome prediction in UK Employment Tribunal (UKET) decisions. Using a corpus of 33,158 individual claims, we predict outcomes from claim texts and LLM-extracted case summaries, evaluating models ranging from interpretable TF-IDF-based classifiers to black-box LLMs. While headline predictive performance figures appear strong, we demonstrate that such performance in LJP systems trained on post-hoc judicial text can be driven by the retrospective nature of the source material. Stratifying the test data by human judgments of leakage reveals that performance increases where outcome-revealing cues are embedded in the narrative. Moreover, a model trained on just the 4% of features identified as leakage achieves high performance, outperforming human experts. These findings substantiate concerns that LJP performance may be exaggerated by linguistic artefacts. Yet this vulnerability is not fatal to the research agenda. Instead, post-hoc judgments might be treated as potentially contaminated texts, requiring active auditing. Retraining models after masking leakage features results in only a negligible reduction in Macro-F1. Hence, while models will opportunistically exploit shortcuts when available, they remain capable of extracting useful predictive signals when these artefacts are removed.

An Intervention-Based Framework for Shortcut Diagnosis in Spoofing Countermeasures

While deepfake audio detection systems achieve high performance in controlled benchmarks, their reliability often diminishes in the wild. Prior work shows that dataset-specific artifacts contribute to this gap. Yet, systematic tools to identify which acoustic properties a model exploits as shortcuts remain limited. We propose an intervention-based diagnostic framework, grounded in a directed graphical model, that formally distinguishes confound-driven shortcut dependencies from legitimate domain shift. We operationalise this through controlled acoustic perturbations targeting non-speech structure, spectral content, and signal energy, complemented by corpus-level distributional analysis. Evaluating XLS-R-300M with RawGAT-ST across ASVspoof challenges datasets, we quantify model sensitivity to specific intervention types. Results reveal that non-speech interventions produce the largest performance shifts, confirming non-speech intervals as a dominant shortcut.

The Moving Eye: Enhancing VLA Spatial Generalization via Hybrid Dynamic Data Collection

Vision-Language-Action (VLA) models have shown remarkable promise in generalized robotic manipulation. However, their spatial generalization remains fragile. We argue that simply increasing the number of viewpoints is insufficient. Models often fall into the trap of Shortcut Learning, latching onto spurious correlations (e.g., fixed relative poses between objects or between the camera and robot base) rather than learning true spatial relationships. In this work, we propose a data-centric solution to enhance VLA spatial generalization. We utilize a dual-arm setup where one arm performs manipulation while the other serves as a mobile environmental camera. We systematically evaluate three data distribution patterns: Fixed, Multi-Fixed, and Moving Views. Our findings reveal that a hybrid strategy, combining continuous camera motion with diverse static viewpoints, yields the best performance by substantially reducing spurious correlations while maintaining training stability. Our experiments demonstrate that this strategy mitigates spurious correlations, enabling VLAs to generalize to unseen camera poses and object configurations where simply adding more static viewpoints fails. Crucially, we reveal that the susceptibility to shortcut learning and the struggle with spatial generalization are universal characteristics shared across diverse architectures. Consequently, all evaluated models (ACT, Diffusion, and VLA models including Pi0 and Gr00t) benefit significantly from our mixed data strategy.

Dataset Biases and Shortcut Learning in Motion-Based AI-Generated Video Detection

The visual quality of AI-generated videos has improved drastically in recent years, making it increasingly difficult for humans to distinguish between real and synthetic media. In this work, we evaluate the robustness and applicability of four state-of-the-art motion-based AI-generated video detectors. We identify significant preprocessing and sampling biases in these methods and demonstrate that they account for a substantial portion of their reported performance. Furthermore, we find that these detectors are highly sensitive to motion patterns specific to their evaluation datasets, where AI-generated videos generally exhibit less inter-frame movement than real videos. We show that for all detectors, performance collapses to near-random levels when evaluated on a dataset that does not contain this motion bias. Additionally, through dataset rebalancing and the application of simple spatial augmentations, we observe severe performance degradation across all evaluated models. In contrast, we find that an existing frequency-based detector maintains strong performance across all evaluated datasets, suggesting that frequency-based approaches may offer a more generalizable path forward for AI-generated video detection. We hope that our work raises awareness towards these vulnerabilities and encourages the development of more representative, unbiased datasets and more robust evaluation protocols.

SHOVIR: A Benchmark for Evaluating Vision Shortcut Learning in Radiology Report Generation

Current evaluation protocols for Vision-Language Models (VLMs) in Radiology Report Generation (RRG) rely on report-level metrics that measure lexical overlap or aggregate clinical correctness. However, such metrics do not test whether individual diagnostic statements stem from the actual pathological evidence visible in the image. This allows models to achieve competitive scores by exploiting learned priors or spurious correlations, a failure mode we refer to as vision shortcut. We introduce SHOVIR, a benchmark for evaluating vision shortcut behavior in RRG. SHOVIR extends two spatially annotated chest X-ray datasets, MIMIC-CXR and PadChest-GR, with per-box CheXpert labels, and defines image-level and disease-level occlusion experiments that contrast baseline performance on clean images against localized, region-specific perturbations. Comparing predictions across these conditions isolates two failure modes at the disease-class level: direct shortcuts, where a finding persists after its visual evidence is removed, and contextual shortcuts, where detection degrades once co-occurring pathologies are occluded despite the target region remaining intact. Benchmarking eight state-of-the-art VLMs, we find that shortcut behavior varies substantially across architectures and datasets. Models achieving the highest baseline report quality do not necessarily rank highest in spatial grounding, revealing that clinically fluent generation can coexist with shallow reliance on visual evidence. These findings expose a blind spot in current RRG evaluation and motivate region-aware assessment protocols.

Breaking Shortcut Learning for Cross-Trial EEG-Guided Target Speech Extraction via Two-Stage Training

Recent end-to-end models for EEG-guided target speech extraction report impressive results, underscoring potential for neuro-steered hearing technologies. However, our analysis reveals that high within-trial performance can be driven by trial-specific EEG structure that acts as shortcuts for target selection, leading to poor generalization on unseen trials. To overcome this gap, we propose TRUST-TSE, a two-stage framework to mitigate shortcut learning. By introducing contrastive pretraining with attended-speaker negative sampling, we encourage the EEG encoder to capture fine-grained EEG--speech alignment while suppressing trial-identity cues. We also employ a confidence-weighted extraction objective based on EEG--source similarity to guide extraction using the learned representations. Experiments on KUL and DTU datasets show that TRUST-TSE outperforms end-to-end baselines under strict cross-trial protocols, addressing a key reliability bottleneck of existing approaches.

Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation

On-policy self-distillation (OPSD) trains a model on its own rollouts and uses a frozen copy to provide dense token-level targets conditioned on a reference target. This works well for LLM reasoning, but a direct extension to multimodal large language models (MLLMs) can create a shortcut: the privileged target may guide tokens mainly based on the text reference target rather than the image. We propose ViGOS, a visually grounded OPSD framework for MLLM post-training. The student first writes a visual description and then reasons toward the final answer. For valid rollouts, an image-only perception teacher supervises the description, while a privileged reasoning teacher supervises the reasoning and final answer on the same student prefix. A reference teacher is used only for invalid rollouts to recover the output format. Across general vision-language, expert reasoning, visual math, spatial grounding, and visual-language-prior benchmarks, ViGOS keeps the main benefits of OPSD and improves image-grounded behavior in shortcut-prone settings.

Assessing Reliability of Symbol Detection in Concept Bottleneck Models

Concept Bottleneck Models (CBMs) are a relevant tool for explainable Artificial Intelligence because they make their predictions through human-interpretable symbols. However, high task accuracy does not guarantee that these symbols are detected faithfully: jointly trained CBMs may encode task-specific shortcuts in the bottleneck, making their explanations unreliable. In this paper, we study concept-detection reliability by swapping independently trained concept detectors and classification heads that share the same symbolic vocabulary. We use the resulting performance degradation, concept-level metrics, and symbol-wise uncertainty estimates to identify concepts that are especially prone to spurious firing. Finally, we propose a reliability-aware training strategy in which a shared concept detector is optimized with multiple classification heads and penalized for relying on globally or instance-wise unreliable symbols. On CUB-200-2011 with full concept supervision, detectors and heads are almost freely interchangeable (swap drop below one accuracy point, relative retention above $99\%$, and no concept detected below chance), whereas on a controlled synthetic task we show that, as the concept-supervision weight is reduced, models keep near-perfect task accuracy while swapped accuracy and agreement with the ground-truth concepts collapse to chance. Our reliability-aware training substantially mitigates this leakage, roughly doubling swap accuracy in the leaky regime.

From Shortcuts to Reasoning: Robust Post-Training of Theory of Mind with Reinforcement Learning

Theory of Mind (ToM) is a must-acquire skill for modern foundation model systems to operate effectively and safely in the real world. Recent works have explored honing ToM via post-training; however, we show that such progress is confounded by a pervasive "shortcut" issue: tasks can reach up to 99% accuracy by simply exploiting spurious causal correlations, leading to a false sense of ToM. Motivated by this, we first develop a framework to systematically examine ToM datasets for shortcuts and provide guidance for future development. We find that questions reducible to pure state tracking, such as "belief," are especially shortcut-prone compared to mind questions, such as "intention," where reasoning beyond tracking is required. Using four shortcut-free datasets across three ToM contexts, we then comprehensively study whether Reinforcement Fine-Tuning with verifiable rewards and explicit reasoning chains, called Thinking-RFT, elevates ToM beyond Supervised Fine-Tuning, or SFT. Our key findings are as follows. First, Thinking-RFT effectively improves ToM in all scenarios, with a 6% improvement over SFT, particularly in complex higher-order reasoning, with a 10% improvement over SFT, and multimodal cases, with a 7% improvement over SFT. It also generalizes notably better to unseen domains and higher-order queries while being more robust to counterfactuals. Second, ToM benefits specifically from the joint effect of reasoning and RL: Thinking-RFT outperforms Non-Thinking-RFT by 7% on average. Third, RFT works by learning to ground its reasoning on anchor cues, such as keywords and state changes, that correspond to causal factors. We believe our study is useful for developing effective and robust ToM post-training datasets and advancing critical ToM capabilities.