Shortcut Learning

Shortcut learning causes deep learning models to rely on non-essential features within the data. However, its formation in deep neural network training still lacks theoretical understanding. In this paper, we provide a formal definition of core and shortcut features and employ evolutionary game theory to analyze the origins of shortcut bias by modeling data samples as players and their corresponding neural tangent features as strategies, assuming the existence of core and shortcut subnetworks. We find that gradient descent (GD) and stochastic gradient descent (SGD) lead to two distinct stochastically stable states, each corresponding to a different strategy. The former primarily optimizes the shortcut subnetwork, while the latter primarily optimizes the core subnetwork. We investigate the influence of these strategies on shortcut bias through a continuous stochastic differential equation, and reveal the impact of data noise and optimization noise on the formation of shortcut bias. In brief, our work employs evolutionary game theory to characterize the dynamics of shortcut bias formation and provides a theoretical view on its mitigation.

Purpose: Vision-language models (VLMs) have shown promising performance in surgical visual question answering (VQA). However, existing surgical VQA datasets often contain linguistic shortcuts, where question phrasing implicitly constrains the answer space. It remains unclear whether reported performance reflects visual understanding or reliance on such linguistic shortcuts. Methods: We introduce SurgCheck, a diagnostic benchmark for quantifying linguistic shortcut reliance in surgical VQA. SurgCheck employs a paired-question design in which each surgical frame is associated with an original question containing entity names and a less-biased counterpart that removes these names while preserving identical visual content and ground-truth answers. The resulting performance gap provides a diagnostic signal of shortcut reliance. To ensure that the less-biased question remains well-defined even without entity names, four grounding cues are incorporated: bounding box, arrow, spatial position, and periphrasis. We evaluate both general-purpose and surgical-specific VLMs under zero-shot and fine-tuned settings on SurgCheck. To evaluate open-ended zero-shot responses, we introduce an LLM-as-a-judge evaluation protocol. Results: Using SurgCheck, we observe consistent performance degradation on less-biased questions across five VLMs, despite identical visual inputs. Text-only ablation reveals minimal performance drops for action and target prediction, indicating that action and target prediction is largely driven by linguistic shortcuts rather than visual reasoning. Conclusion: SurgCheck provides a controlled diagnostic framework that exposes failure modes masked by linguistic bias in existing surgical VQA benchmarks. Our findings demonstrate that strong benchmark performance does not necessarily imply faithful visual understanding, underscoring the need for bias-aware evaluation in surgical VQA.

Although reinforcement learning (RL) has significantly advanced reasoning capabilities in large multimodal language models (MLLMs), its efficacy remains limited for lightweight models essential for edge deployments. To address this issue, we leverage causal analysis and experiment to reveal the underlying phenomenon of perceptual bias, demonstrating that RL-based fine-tuning compels lightweight models to preferentially adopt perceptual shortcuts induced by data biases, rather than developing genuine reasoning abilities. Motivated by this insight, we propose VideoThinker, a causal-inspired framework that cultivates robust reasoning in lightweight models through a two-stage debiasing process. First, the Bias Aware Training stage forges a dedicated "bias model" to embody these shortcut behaviors. Then, the Causal Debiasing Policy Optimization (CDPO) algorithm fine-tunes the primary model, employing an innovative repulsive objective to actively push it away from the bias model's flawed logic while simultaneously pulling it toward correct, generalizable solutions. Our model, VideoThinker-R1, establishes a new state-of-the-art in video reasoning efficiency. For same-scale comparison, requiring no Supervised Fine-Tuning (SFT) and using only 1 of the training data for RL, it surpasses VideoRFT-3B with a 3.2% average gain on widely-used benchmarks and a 7% lead on VideoMME. For cross-scale comparison, it outperforms the larger Video-UTR-7B model on multiple benchmarks, including a 2.1% gain on MVBench and a 3.8% gain on TempCompass. Code is available at https://github.com/falonss703/VideoThinker.

Prior work on node classification has shown that Graph Neural Networks (GNNs) can learn representations that transfer across graphs, when underlying graph properties are shared. For a fixed graph, one would then expect GNNs trained for link prediction to learn a representation consistent with that learnt for node classification. We show this intuition does not hold in the general case. Instead, we find popular link prediction models can learn a trivial mini-batch dependent heuristic, enabled by batch-normalisation layers, to solve the edge classification task. When correcting for this, we observe increased alignment of the network representation with node-class relevant features, suggesting the network has learnt a graph representation that better aligns with the underlying graph's properties. Our findings suggest that standard link prediction training may be leading us to overestimate link predictors' ability to learn a generalised representation of a graph that is consistent across tasks.

Monitoring frying oil degradation is critical for food safety, yet current practice relies on destructive wet-chemistry assays that provide no spatial information and are unsuitable for real-time use. We identify a fundamental obstacle in thermal-image-based inspection, the camera-fingerprint shortcut, whereby models memorize sensor-specific noise and thermal bias instead of learning oxidation chemistry, collapsing under video-disjoint evaluation. We propose FryNet, a dual-stream RGB-thermal framework that jointly performs oil-region segmentation, serviceability classification, and regression of four chemical oxidation indices (PV, p-AV, Totox, temperature) in a single forward pass. A ThermalMiT-B2 backbone with channel and spatial attention extracts thermal features, while an RGB-MAE Encoder learns chemically grounded representations via masked autoencoding and chemical alignment. Dual-Encoder DANN adversarially regularizes both streams against video identity via Gradient Reversal Layers, and FiLM fusion bridges thermal structure with RGB chemical context. On 7,226 paired frames across 28 frying videos, FryNet achieves 98.97% mIoU, 100% classification accuracy, and 2.32 mean regression MAE, outperforming all seven baselines.

Recent Vision-Language-Action (VLA) models report impressive success rates on standard robotic benchmarks, fueling optimism about general-purpose physical intelligence. However, recent evidence suggests a systematic misalignment between standard benchmark success and true embodied reasoning, raising the question of whether these high scores reflect genuine cognitive capability. To address this gap, we introduce BeTTER, a diagnostic Benchmark for Testing True Embodied Reasoning in robotic policies. BeTTER applies targeted causal interventions (e.g., spatial layout shifts, temporal extrapolation) while enforcing kinematic isolation to explicitly decouple high-level reasoning failures from low-level execution limits. Through systematic evaluation, we reveal that state-of-the-art VLAs catastrophically fail in dynamic scenarios, exhibiting severe lexical-kinematic shortcuts, behavioral inertia, and semantic feature collapse. Crucially, our mechanistic analysis traces these symptoms to fundamental architectural bottlenecks - such as capacity compression and myopic downsampling - which systematically degrade the model's foundational semantic representation. We demonstrate that highly static evaluation protocols effectively mask this degradation by allowing optimization to overfit to sensorimotor priors. Supported by real-world robotic validation, our findings confirm that this representational breakdown is not a simulation artifact, highlighting the critical need for future VLA paradigms to resolve the structural tension between high-frequency control and high-level reasoning.

Multimodal Reasoning Models (MRMs) leveraging Chain-of-Thought (CoT) based thinking have revolutionized mathematical and logical problem-solving. However, we show that this paradigm struggles with generalized spatial intelligence. We perform a comprehensive evaluation of seventeen models across thirteen spatial benchmarks and identify a critical gap: CoT prompting consistently degrades performance in visual spatial reasoning. Furthermore, through a novel No-Image++ ablation, we demonstrate that MRMs and CoT prompted MLMs suffer from severe shortcut learning, and hallucinate visual details from textual priors even when the image is absent. These findings challenge the efficacy of text-only CoT for spatial tasks and underscore the need for vision-centric reasoning paradigms.

Pretrained text encoders are prone to shortcut learning, relying on token-label correlations that fail once the distribution shifts in deployment. Existing shortcut mitigation methods mainly operate at training time and assume access to training data, training dynamics, or shortcut annotations, which are hardly available during deployment, where only the converged model remains. We show that this model alone suffices to mitigate shortcuts during deployment: a biased model internalizes a signal of its learned shortcuts that can be captured via unsupervised gradient-based attribution. We further prove that deployment-time mitigation is information-theoretically upper-bounded by training-time mitigation. Nevertheless, exploiting this gradient signal, our proposed unsupervised deployment-time shortcut mitigation framework for pretrained text encoders, Shortcut Guardrail, recovers substantial performance under shortcut distribution shift, matching or outperforming training-time baselines across sentiment classification, toxicity detection, and natural language inference.

Deep Neural Networks are highly susceptible to shortcut learning, frequently memorizing low-dimensional spurious correlations instead of underlying causal mechanisms. This phenomenon not only degrades out-of-distribution robustness but also induces severe demographic biases in sensitive applications. In this paper, we propose a geometric a priori methodology to mitigate shortcut learning. By deploying a zero-hidden-layer ($N=1$) Topological Auditor, we mathematically isolate features that monopolize the gradient without human intervention. We empirically demonstrate a Capacity Phase Transition: once linear shortcuts are pruned, networks are forced to utilize higher geometric capacity ($N \geq 16$) to curve the decision boundary and learn ethical representations. Our approach outperforms L1 Regularization -- which collapses into demographic bias -- and operates at a fraction of the computational cost of post-hoc methods like Just Train Twice (JTT), successfully reducing counterfactual gender vulnerability from 21.18% to 7.66%.

Scientific claim verification, the task of determining whether claims are entailed by scientific evidence, is fundamental to establishing discoveries in evidence while preventing misinformation. This process involves evaluating each asserted constraint against validated evidence. Under the Closed-World Assumption (CWA), a claim is accepted if and only if all asserted constraints are positively supported. We show that existing verification benchmarks cannot distinguish models enforcing this standard from models applying a simpler shortcut called salient-constraint checking, which applies CWA's rejection criterion only to the most salient constraint and accepts when that constraint is supported. Because existing benchmarks construct infeasible claims by perturbing a single salient element they are insufficient at distinguishing between rigorous claim verification and simple salient-constraint reliance. To separate the two, we construct compositionally infeasible claims where the salient constraint is supported but a non-salient constraint is contradicted. Across model families and modalities, models that otherwise saturate existing benchmarks consistently over-accept these claims, confirming the prevalence of such shortcut reasoning. Via model context interventions, we show that different models and prompting strategies occupy distinct positions on a shared ROC curve, indicating that the gap between model families reflects differences in verification threshold rather than underlying reasoning ability, and that the compositional inference bottleneck is a structural property of current verification behavior that strategy guidance alone cannot overcome.

To extend the reinforcement learning post-training paradigm to omni-modal models for concurrently bolstering video-audio understanding and collaborative reasoning, we propose OmniJigsaw, a generic self-supervised framework built upon a temporal reordering proxy task. Centered on the chronological reconstruction of shuffled audio-visual clips, this paradigm strategically orchestrates visual and auditory signals to compel cross-modal integration through three distinct strategies: Joint Modality Integration, Sample-level Modality Selection, and Clip-level Modality Masking. Recognizing that the efficacy of such proxy tasks is fundamentally tied to puzzle quality, we design a two-stage coarse-to-fine data filtering pipeline, which facilitates the efficient adaptation of OmniJigsaw to massive unannotated omni-modal data. Our analysis reveals a ``bi-modal shortcut phenomenon'' in joint modality integration and demonstrates that fine-grained clip-level modality masking mitigates this issue while outperforming sample-level modality selection. Extensive evaluations on 15 benchmarks show substantial gains in video, audio, and collaborative reasoning, validating OmniJigsaw as a scalable paradigm for self-supervised omni-modal learning.

Stain variability is a pervasive source of distribution shift and potential shortcut learning in renal pathology AI. We ask whether lupus nephritis glomerular lesion classifiers exploit stain as a shortcut, and how to mitigate such bias without stain or site labels. We curate a multi-center, multi-stain dataset of 9,674 glomerular patches (224$\times$224) from 365 WSIs across three centers and four stains (PAS, H&E, Jones, Trichrome), labeled as proliferative vs. non-proliferative. We evaluate Bayesian CNN and ViT backbones with Monte Carlo dropout in three settings: (1) stain-only classification; (2) a dual-head model jointly predicting lesion and stain with supervised stain loss; and (3) a dual-head model with label-free stain regularization via entropy maximization on the stain head. In (1), stain identity is trivially learnable, confirming a strong candidate shortcut. In (2), varying the strength and sign of stain supervision strongly modulates stain performance but leaves lesion metrics essentially unchanged, indicating no measurable stain-driven shortcut learning on this multi-stain, multi-center dataset, while overly adversarial stain penalties inflate predictive uncertainty. In (3), entropy-based regularization holds stain predictions near chance without degrading lesion accuracy or calibration. Overall, a carefully curated multi-stain dataset can be inherently robust to stain shortcuts, and a Bayesian dual-head architecture with label-free entropy regularization offers a simple, deployment-friendly safeguard against potential stain-related drift in glomerular AI.

Jaguar re-identification (re-ID) from citizen-science imagery can look strong on standard retrieval metrics while still relying on the wrong evidence, such as background context or silhouette shape, instead of the coat pattern that defines identity. We introduce a diagnostic framework for wildlife re-ID with two axes: a leakage-controlled context ratio, background/foreground, computed from inpainted background-only versus foreground-only images, and a laterality diagnostic based on cross-flank retrieval and mirror self-similarity. To make these diagnostics measurable, we curate a Pantanal jaguar benchmark with per-pixel segmentation masks and an identity-balanced evaluation protocol. We then use representative mitigation families, ArcFace fine-tuning, anti-symmetry regularization, and Lorentz hyperbolic embeddings, as case studies under the same evaluation lens. The goal is not only to ask which model ranks best, but also what visual evidence it uses to do so.

Deep Neural Networks (DNNs) are increasingly utilized in high-stakes domains like medical diagnostics and autonomous driving where model reliability is critical. However, the research landscape for ensuring this reliability is terminologically fractured across communities that pursue the same goal of ensuring models rely on causally relevant features rather than confounding signals. While frameworks such as distributionally robust optimization (DRO), invariant risk minimization (IRM), shortcut learning, simplicity bias, and the Clever Hans effect all address model failure due to spurious correlations, researchers typically only reference work within their own domains. This reproducibility study unifies these perspectives through a comparative analysis of correction methods under challenging constraints like limited data availability and severe subgroup imbalance. We evaluate recently proposed correction methods based on explainable artificial intelligence (XAI) techniques alongside popular non-XAI baselines using both synthetic and real-world datasets. Findings show that XAI-based methods generally outperform non-XAI approaches, with Counterfactual Knowledge Distillation (CFKD) proving most consistently effective at improving generalization. Our experiments also reveal that the practical application of many methods is hindered by a dependency on group labels, as manual annotation is often infeasible and automated tools like Spectral Relevance Analysis (SpRAy) struggle with complex features and severe imbalance. Furthermore, the scarcity of minority group samples in validation sets renders model selection and hyperparameter tuning unreliable, posing a significant obstacle to the deployment of robust and trustworthy models in safety-critical areas.

Current synthetic data pipelines for computer vision generate images without diagnosing what the downstream model actually needs. This open-loop paradigm treats synthetic data as cheap real data, randomly sampling the generator's output space and hoping to cover the model's failure modes. We argue this fundamentally misuses synthetic data's unique property: the controllable, independent variation of scene factors.Drawing on the statistical theory of Design of Experiments (DoE), we propose Synthetic Designed Experiments for Representational Sufficiency (SDRS). SDRS treats the downstream model as a black-box system and the synthetic generator as an experimental apparatus. Using fractional factorial designs, SDRS efficiently audits a model's factor-sensitivity profile via ANOVA decomposition. It classifies failures into two actionable types: Type I gaps (coverage failures on underrepresented factor levels) and Type II gaps (reliance on spurious nuisance dependencies). The audit then prescribes targeted synthetic data to address each gap type. We validate SDRS on three experiments: (1) a controlled diagnostic on dSprites with planted biases, where the audit correctly identifies both gap types and targeted data improves accuracy from 49.9% to 79.0%; (2) a dense segmentation task on procedural scenes, where detecting background-complexity shortcuts and applying targeted data improves mIoU from 0.948 to 0.998; and (3) an entanglement detection experiment showing that the ANOVA audit identifies cross-factor contamination in imperfect generators. Finally, we show that per-factor invariance penalties can transfer sensitivity between factors, identifying an open problem for representation-level correction.

Deep neural networks exhibit a simplicity bias, a well-documented tendency to favor simple functions over complex ones. In this work, we cast new light on this phenomenon through the lens of the Minimum Description Length principle, formalizing supervised learning as a problem of optimal two-part lossless compression. Our theory explains how simplicity bias governs feature selection in neural networks through a fundamental trade-off between model complexity (the cost of describing the hypothesis) and predictive power (the cost of describing the data). Our framework predicts that as the amount of available training data increases, learners transition through qualitatively different features -- from simple spurious shortcuts to complex features -- only when the reduction in data encoding cost justifies the increased model complexity. Consequently, we identify distinct data regimes where increasing data promotes robustness by ruling out trivial shortcuts, and conversely, regimes where limiting data can act as a form of complexity-based regularization, preventing the learning of unreliable complex environmental cues. We validate our theory on a semi-synthetic benchmark showing that the feature selection of neural networks follows the same trajectory of solutions as optimal two-part compressors.

Out-of-Distribution (OOD) generalization has become a primary metric for evaluating image analysis systems. Since deep learning models tend to capture domain-specific context, they often develop shortcut dependencies on these non-causal features, leading to inconsistent performance across different data sources. Current techniques, such as invariance learning, attempt to mitigate this. However, they struggle to isolate highly mixed features within deep latent spaces. This limitation prevents them from fully resolving the shortcut learning problem.In this paper, we propose Hierarchical Causal Dropout (HCD), a method that uses channel-level causal masks to enforce feature sparsity. This approach allows the model to separate causal features from spurious ones, effectively performing a causal intervention at the representation level. The training is guided by a Matrix-based Mutual Information (MMI) objective to minimize the mutual information between latent features and domain labels, while simultaneously maximizing the information shared with class labels.To ensure stability, we incorporate a StyleMix-driven VICReg module, which prevents the masks from accidentally filtering out essential causal data. Experimental results on OOD benchmarks show that HCD performs better than existing top-tier methods.

Automatic depression detection from doctor-patient conversations has gained momentum thanks to the availability of public corpora and advances in language modeling. However, interpretability remains limited: strong performance is often reported without revealing what drives predictions. We analyze three datasets: ANDROIDS, DAIC-WOZ, E-DAIC and identify a systematic bias from interviewer prompts in semi-structured interviews. Models trained on interviewer turns exploit fixed prompts and positions to distinguish depressed from control subjects, often achieving high classification scores without using participant language. Restricting models to participant utterances distributes decision evidence more broadly and reflects genuine linguistic cues. While semi-structured protocols ensure consistency, including interviewer prompts inflates performance by leveraging script artifacts. Our results highlight a cross-dataset, architecture-agnostic bias and emphasize the need for analyses that localize decision evidence by time and speaker to ensure models learn from participants' language.

Physical knot classification is a challenging fine-grained recognition task in which the intended discriminative cue is rope crossing structure; however, high closed-set accuracy may still arise from low-level appearance shortcuts rather than genuine topological understanding. In this work, we introduce dataset (1,440 images, 10 classes), which trains models on loosely tied knots and evaluates them on tightly dressed configurations to probe whether structure-guided training yields topology-specific gains. We demonstrate that topological distance successfully predicts residual inter-class confusion across multiple backbone architectures, validating the utility of our topology-aware evaluation framework. Furthermore, we propose topology-aware centroid alignment (TACA) and an auxiliary crossing-number prediction objective as two complementary forms of structural supervision. Notably, Swin-T with TACA achieves a consistent positive specificity gain (Delta_spec = +1.18 pp) across all random seeds under the canonical protocol, and auxiliary crossing-number prediction exhibits robust performance across data regimes without the real-versus-random reversal observed for centroid alignment. Causal probes reveal that background changes alone flip 17-32% of predictions and phone-photo accuracy drops by 58-69 percentage points, underscoring that appearance bias remains the principal obstacle to deployment. These results collectively demonstrate that our diagnostic workflow provides a principled and practical tool for evaluating whether a hand-crafted structural prior delivers genuine task-relevant benefit beyond generic regularization.

The use of ML in cybersecurity has long been impaired by generalization issues: Models that work well in controlled scenarios fail to maintain performance in production. The root cause often lies in ML algorithms learning superficial patterns (shortcuts) rather than underlying cybersecurity concepts. We investigate contrastive multi-modal learning as a first step towards improving ML performance in cybersecurity tasks. We aim at transferring knowledge from data-rich modalities, such as text, to data-scarce modalities, such as payloads. We set up a case study on threat classification and propose a two-stage multi-modal contrastive learning framework that uses textual vulnerability descriptions to guide payload classification. First, we construct a semantically meaningful embedding space using contrastive learning on descriptions. Then, we align payloads to this space, transferring knowledge from text to payloads. We evaluate the approach on a large-scale private dataset and a synthetic benchmark built from public CVE descriptions and LLM-generated payloads. The methodology appears to reduce shortcut learning over baselines on both benchmarks. We release our synthetic benchmark and source code as open source.

Planning-oriented end-to-end driving models show great promise, yet they fundamentally learn statistical correlations instead of true causal relationships. This vulnerability leads to causal confusion, where models exploit dataset biases as shortcuts, critically harming their reliability and safety in complex scenarios. To address this, we introduce CausalVAD, a de-confounding training framework that leverages causal intervention. At its core, we design the sparse causal intervention scheme (SCIS), a lightweight, plug-and-play module to instantiate the backdoor adjustment theory in neural networks. SCIS constructs a dictionary of prototypes representing latent driving contexts. It then uses this dictionary to intervene on the model's sparse vectorized queries. This step actively eliminates spurious associations induced by confounders, thereby eliminating spurious factors from the representations for downstream tasks. Extensive experiments on benchmarks like nuScenes show CausalVAD achieves state-of-the-art planning accuracy and safety. Furthermore, our method demonstrates superior robustness against both data bias and noisy scenarios configured to induce causal confusion.

Despite recent advances in face recognition, robust performance remains challenging under large variations in age, pose, and occlusion. A common strategy to address these issues is to guide representation learning with auxiliary supervision from facial attributes, encouraging the visual encoder to focus on identity-relevant regions. However, existing approaches typically rely on heterogeneous and fixed sets of attributes, implicitly assuming equal relevance across attributes. This assumption is suboptimal, as different attributes exhibit varying discriminative power for identity recognition, and some may even introduce harmful biases. In this paper, we propose an attribute-aware face recognition architecture that supervises the learning of facial embeddings using identity class labels, identity-relevant facial attributes, and non-identity-related attributes. Facial attributes are organized into interpretable groups, making it possible to decompose and analyze their individual contributions in a human-understandable manner. Experiments on standard face verification benchmarks demonstrate that joint learning of identity and facial attributes improves the discriminability of face embeddings with two major conclusions: (i) using identity-relevant subsets of facial attributes consistently outperforms supervision with a broader attribute set, and (ii) explicitly forcing embeddings to unlearn non-identity-related attributes yields further performance gains compared to leaving such attributes unsupervised. Additionally, our method serves as a diagnostic tool for assessing the trustworthiness of face recognition encoders by allowing for the measurement of accuracy gains with suppression of non-identity-relevant attributes, with such gains suggesting shortcut learning from redundant attributes associated with each identity.

Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.

The growing realism of generative models has blurred the boundary between real and synthetic content, posing significant challenges to reliable AI-generated image detection. Although large-scale pre-trained Vision Foundation Models have advanced detection capability, their generalization to images from unseen generation pipelines remains inadequate. In this paper, we identify, for the first time, a key failure mechanism, termed semantic fallback, wherein forensic fine-tuning fails to fully reshape the representation space. Consequently, the resulting representations remain organized along high-level semantic structures rather than manipulation-specific forensic cues. Building on this insight, we propose a Geometric Semantic Decoupling (GSD) framework, which explicitly suppresses semantically dominant directions, thereby promoting invariant forensic representations. Specifically, GSD leverages a frozen CLIP encoder to estimate the dominant semantic subspace via Singular Value Decomposition (SVD). It then suppresses the semantic components through a geometry-constrained formulation with the suppression strength adaptively modulated across samples and layers. We further introduce a mini-batch SVD approximation strategy that amortizes subspace estimation, achieving over a $15 \times$ reduction in computational overhead while preserving effectiveness. Finally, considering practical scenarios spanning both large-scale and online evaluation, we develop three inference protocols, batch, per-sample, and reference-based inference, and demonstrate that they induce consistent semantic decoupling, yielding a stable forgery-oriented feature manifold.

Neural networks often rely on spurious shortcuts for many epochs before discovering structured representations. However, the mechanism governing when this transition occurs and whether its timing can be predicted remains unclear. Prior work shows that gradient descent converges to low norm solutions and that neural networks exhibit simplicity bias, but neither explains the timescale of the transition from shortcut features to structured representations. We introduce the Norm-Hierarchy Transition (NHT) framework, which explains delayed representation learning as the slow traversal of a hierarchy of parameter norms during regularized optimization. When multiple interpolating solutions exist with different norms, weight decay gradually moves the model from high norm shortcut solutions toward lower norm structured representations. We derive a tight bound showing that the transition delay grows logarithmically with the ratio between shortcut and structured norms. Experiments on modular arithmetic, CIFAR-10 with spurious features, CelebA, and Waterbirds support the predictions of the framework. The results suggest that grokking, shortcut learning, and delayed feature discovery arise from a common mechanism based on norm hierarchy traversal during training.

Accuracy-based evaluation cannot reliably distinguish genuine generalization from shortcuts like memorization, leakage, or brittle heuristics, especially in small-data regimes. In this position paper, we argue for mechanism-aware evaluation that combines task-relevant symbolic rules with mechanistic interpretability, yielding algorithmic pass/fail scores that show exactly where models generalize versus exploit patterns. We demonstrate this on NL-to-SQL by training two identical architectures under different conditions: one without schema information (forcing memorization), one with schema (enabling grounding). Standard evaluation shows the memorization model achieves 94% field-name accuracy on unseen data, falsely suggesting competence. Our symbolic-mechanistic evaluation reveals this model violates core schema generalization rules, a failure invisible to accuracy metrics.

Deep learning models can identify racial identity with high accuracy from chest X-ray (CXR) recordings. Thus, there is widespread concern about the potential for racial shortcut learning, where a model inadvertently learns to systematically bias its diagnostic predictions as a function of racial identity. Such racial biases threaten healthcare equity and model reliability, as models may systematically misdiagnose certain demographic groups. Since racial shortcuts are diffuse - non-localized and distributed throughout the whole CXR recording - image preprocessing methods may influence racial shortcut learning, yet the potential of such methods for reducing biases remains underexplored. Here, we investigate the effects of image preprocessing methods including lung masking, lung cropping, and Contrast Limited Adaptive Histogram Equalization (CLAHE). These approaches aim to suppress spurious cues encoding racial information while preserving diagnostic accuracy. Our experiments reveal that simple bounding box-based lung cropping can be an effective strategy for reducing racial shortcut learning while maintaining diagnostic model performance, bypassing frequently postulated fairness-accuracy trade-offs.

Latent reasoning has been recently proposed as a reasoning paradigm and performs multi-step reasoning through generating steps in the latent space instead of the textual space. This paradigm enables reasoning beyond discrete language tokens by performing multi-step computation in continuous latent spaces. Although there have been numerous studies focusing on improving the performance of latent reasoning, its internal mechanisms remain not fully investigated. In this work, we conduct a comprehensive analysis of latent reasoning methods to better understand the role and behavior of latent representation in the process. We identify two key issues across latent reasoning methods with different levels of supervision. First, we observe pervasive shortcut behavior, where they achieve high accuracy without relying on latent reasoning. Second, we examine the hypothesis that latent reasoning supports BFS-like exploration in latent space, and find that while latent representations can encode multiple possibilities, the reasoning process does not faithfully implement structured search, but instead exhibits implicit pruning and compression. Finally, our findings reveal a trade-off associated with supervision strength: stronger supervision mitigates shortcut behavior but restricts the ability of latent representations to maintain diverse hypotheses, whereas weaker supervision allows richer latent representations at the cost of increased shortcut behavior.

Vision Transformers (ViTs), when pre-trained on large-scale data, provide general-purpose representations for diverse downstream tasks. However, artifacts in ViTs are widely observed across different supervision paradigms and downstream tasks. Through systematic analysis of artifacts in ViTs, we find that their fundamental mechanisms have yet to be sufficiently elucidated. In this paper, through systematic analysis, we conclude that these artifacts originate from a lazy aggregation behavior: ViT uses semantically irrelevant background patches as shortcuts to represent global semantics, driven by global attention and Coarse-grained semantic supervision. Our solution selectively integrates patch features into the CLS token, reducing the influence of background-dominated shortcuts and consistently improving performance across 12 benchmarks under label-, text-, and self-supervision. We hope this work offers a new perspective on ViT behavior.

Machine unlearning, which enables a model to forget specific data, is crucial for ensuring data privacy and model reliability. However, its effectiveness can be severely undermined in real-world scenarios where models learn unintended biases from spurious correlations within the data. This paper investigates the unique challenges of unlearning from such biased models. We identify a novel phenomenon we term shortcut unlearning," where models exhibit an easy to learn, yet hard to forget" tendency. Specifically, models struggle to forget easily-learned, bias-aligned samples; instead of forgetting the class attribute, they unlearn the bias attribute, which can paradoxically improve accuracy on the class intended to be forgotten. To address this, we propose CUPID, a new unlearning framework inspired by the observation that samples with different biases exhibit distinct loss landscape sharpness. Our method first partitions the forget set into causal- and bias-approximated subsets based on sample sharpness, then disentangles model parameters into causal and bias pathways, and finally performs a targeted update by routing refined causal and bias gradients to their respective pathways. Extensive experiments on biased datasets including Waterbirds, BAR, and Biased NICO++ demonstrate that our method achieves state-of-the-art forgetting performance and effectively mitigates the shortcut unlearning problem.