Accepted Papers

Spurious correlations in the data, where multiple cues are predictive of the target labels, often lead to a phenomenon known as shortcut learning, where a model relies on erroneous, easy-to-learn cues while ignoring reliable ones. In this work, we propose $DiffDiv$ an ensemble diversification framework exploiting Diffusion Probabilistic Models (DPMs) to mitigate this form of bias. We show that at particular training intervals, DPMs can generate images with novel feature combinations, even when trained on samples displaying correlated input features. We leverage this crucial property to generate synthetic counterfactuals to increase model diversity via ensemble disagreement. We show that DPM-guided diversification is sufficient to remove dependence on shortcut cues, without a need for additional supervised signals. We further empirically quantify its efficacy on several diversification objectives, and finally show improved generalization and diversification on par with prior work that relies on auxiliary data collection.

Out-of-distribution (OOD) robustness is a desired property of computer vision models. Improving model robustness requires high-quality signals from robustness benchmarks to quantify progress. While various benchmark datasets such as ImageNet-C were proposed in the ImageNet era, most ImageNet-C corruption types are no longer OOD relative to today's large datasets scraped from the web, which already contain common corruptions such as blur or JPEG compression artifacts. Consequently, these standard benchmarks are no longer well-suited for evaluating OOD robustness in the era of web-scale datasets. Indeed, recent models show saturating scores on ImageNet-era OOD benchmarks, indicating that it is unclear whether models trained on web-scale datasets truly become better at OOD generalization or whether they have simply been exposed to the test distortions during training. To address this, we here introduce LAION-C as a benchmark alternative for ImageNet-C. LAION-C consists of six novel distortion types specifically designed to be OOD, even for web-scale datasets such as LAION. In a comprehensive evaluation of state-of-the-art models, we find that the LAION-C dataset poses significant challenges to contemporary models, including MLLMs such as Gemini and GPT-4o. We additionally conducted a psychophysical experiment to evaluate the difficulty of our proposed corruptions for human observers, enabling a comparison of models to lab-quality human robustness data. We observe a paradigm shift in OOD generalization: from humans outperforming models, to the best models now matching or outperforming the best human observers.

Learning models have been shown to rely on spurious correlations between non-predictive features and the associated labels in the training data, with negative implications on robustness, bias and fairness. In this work, we provide a statistical characterization of this phenomenon for high-dimensional regression, when the data contains a predictive *core* feature $x$ and a *spurious* feature $y$. Specifically, we quantify the amount of spurious correlations $\mathcal C$ learned via linear regression, in terms of the data covariance and the strength $\lambda$ of the ridge regularization. As a consequence, we first capture the simplicity of $y$ through the spectrum of its covariance, and its correlation with $x$ through the Schur complement of the full data covariance. Next, we prove a trade-off between $\mathcal C$ and the in-distribution test loss $\mathcal L$, by showing that the value of $\lambda$ that minimizes $\mathcal L$ lies in an interval where $\mathcal C$ is increasing. Finally, we investigate the effects of over-parameterization via the random features model, by showing its equivalence to regularized linear regression. Our theoretical results are supported by numerical experiments on Gaussian, Color-MNIST, and CIFAR-10 datasets.

We propose a general and unifying framework for causal Imitation Learning (IL) with hidden confounders that subsumes several existing settings. Our framework accounts for two types of hidden confounders: (a) those observed by the expert but not the imitator, and (b) confounding noise hidden to both. By leveraging trajectory histories as instruments, we reformulate causal IL into Conditional Moment Restrictions (CMRs). We propose DML-IL, an algorithm that solves these CMRs via instrumental variable regression, and upper bound its imitation gap. Empirical evaluation on continuous state-action environments, including Mujoco tasks, shows that DML-IL outperforms state-of-the-art causal IL methods.

Machine learning models often learn unintended shortcuts (spurious correlations) that do not reflect the true causal structure of a task and thus degrade dramatically under subpopulation shift. This problem becomes especially severe in high-stakes domains where the cost of relying on misaligned shortcuts is prohibitive. To address this challenge, concept bottlenecks explicitly factor predictions into high-level concepts and a simple decision layer, enabling experts to diagnose whether learned concepts align with their domain knowledge. Yet, simply removing undesirable concepts after training is insufficient to prevent shortcuts when the concept encoder is incomplete or entangled. In this work, we propose *CBDebug*, a novel framework to debug concept bottlenecks for robustness under subpopulation shift. First, a domain expert identifies and removes spurious concepts using model explanations (the *Removal* step). Then, leveraging this human feedback, we disentangle or replace the removed shortcuts by retraining on a rebalanced dataset based on the causal graph (the *Retraining* step). Empirically, *CBDebug* significantly outperforms existing concept-based methods. Overall, our work demonstrates how expert-guided debugging of concept bottlenecks can achieve interpretability and robustness, promoting alignment of a model’s internal reasoning with how humans reason.

Neural networks are vulnerable to privacy attacks aimed at stealing sensitive data. The risks are amplified in real-world scenario when models are trained on limited and biased data. In this work, we investigate the impact of spurious correlation bias on privacy vulnerability. We introduce _spurious privacy leakage_, a phenomenon where spurious groups are more vulnerable to privacy attacks compared to other groups. Through empirical analysis, we counterintuitively demonstrate that reducing spurious correlation fails to address the privacy disparity between groups. This leads us to introduce a new perspective on privacy disparity based on data memorization. We show that mitigating spurious correlation does not reduce the degree of data memorization, and therefore, neither the privacy risks. Our findings highlight the need to rethink privacy with spurious learning.

The goal of clustering is to group similar data points together. Deep clustering enhances this process by using neural networks for inferring better data representations through a three-stage approach: pre-training for initial feature learning, deep clustering to structure the latent space, and self-labeling to iteratively refine both representations and cluster assignments. Ever since its inception, self-labeling has been a crucial element for reaching state-of-the-art performance in deep clustering. The samples for the self-labeling phase are obtained by setting a confidence threshold for the network’s predictions and only using samples that exceed this threshold for further training. This often improves clustering performance but relies on training with noisy, self-constructed labels (pseudo-labels). As the model iteratively retrains on its own pseudo-labels, the certainty of its predictions tends to rise, increasing its confidence over time. The increasing confidence leads to a growing number of training samples also including more and more samples assigned to the wrong cluster, which can limit performance. Particularly, the model's initially learned biases are amplified by relying on easily learned but ultimately misleading patterns in pseudo-labels, hampering generalization. In this paper, we propose ReSL, a framework that unites Resets with Self-Labeling. We demonstrate that employing weight-reset techniques during self-labeling increases clustering performance and improves generalization. Our findings address limitations of self-labeling and provide a foundation for future research in developing more robust approaches.

Large language models (LLMs) demonstrate remarkable performance on many NLP tasks, yet often exhibit order dependence: simply reordering semantically identical tokens (e.g., answer choices in multiple-choice questions) can lead to inconsistent predictions. Recent work proposes Set-Based Prompting (SBP) as a way to remove order information from designated token subsets, thereby mitigating positional biases. However, applying SBP on base models induces an out-of-distribution input format, which can degrade in-distribution performance. We introduce a fine-tuning strategy that integrates SBP into the training process, “pulling” these set-formatted prompts closer to the model’s training manifold. We show that SBP can be incorporated into a model via fine-tuning. Our experiments on in-distribution (MMLU) and out-of-distribution (CSQA, ARC Challenge) multiple-choice tasks show that SBP fine-tuning significantly improves accuracy and robustness to answer-order permutations, all while preserving broader language modeling capabilities. We discuss the broader implications of order-invariant modeling and outline future directions for building fairer, more consistent LLMs.

Over the last decade, deep learning models have been widely used for automatic feature extraction and classification in various Brain-Computer Interface (BCI) tasks. However, their performance and generalization capabilities are often not adequately assessed, as these models are frequently trained and tested under flawed setups and / or influenced by spurious correlations. Recently, these limitations have also been observed in the training and evaluation of Large Brainwave Foundation Models (LBMs). In this work, we employ causal reasoning and careful consideration for task-discriminative artifacts in various EEG datasets covering diverse BCI paradigms and propose a benchmarking protocol to properly evaluate the decoding performance and generalization capabilities of LBMs. Utilising a subject-independent cross-validation approach for each curated benchmark dataset, we showcase that LBMs achieve marginal performance gains over conventional deep learning baselines.

A plethora of real-world scientific investigations is waiting to scale with the support of trustworthy predictive models that can reduce the need for costly data annotations. We focus on causal inferences on a target experiment with unlabeled factual outcomes, retrieved by a predictive model fine-tuned on a labeled similar experiment. First, we show that factual outcome estimation via Empirical Risk Minimization (ERM) may fail to yield valid causal inferences on the target population, even in a randomized controlled experiment and infinite training samples. Then, we propose to leverage the observed experimental settings during training to empower generalization to downstream interventional investigations, ``Causal Lifting'' the predictive model. We propose Deconfounded Empirical Risk Minimization (DERM), a new simple learning procedure minimizing the risk over a fictitious target population, preventing potential confounding effects. We validate our method on both synthetic and real-world scientific data. Notably, for the first time, we zero-shot generalize causal inferences on ISTAnt dataset (without annotation) by causal lifting a predictive model on our experiment variant.

Shortcuts, spurious patterns that perform well only on the training distribution, pose a major challenge to deep network reliability (Geirhos et al., 2020). In this work, we investigate the layer-wise impact of image shortcuts on learned features. First, we propose an experiment design that introduces artificial shortcut-inducing skews during training, enabling a counterfactual analysis of how different layers contribute to shortcut-related accuracy degradation. Next, we use our method to study the effects of a patch-like skew on CNNs trained on CIFAR-10 and CIFAR-100. Our analysis reveals that different types of skews affect networks layers differently: class-universal skews (affecting all instances of a target class) and class-specific skews (affecting only one class) impact deeper layers more than non-universal and non-specific skews, respectively. Additionally, we identify the forgetting of shortcut-free features as a key mechanism behind accuracy drop for our class of skews, indicating the potential role of simplicity bias (Shah et al., 2020) and excessive regularization (Sagawa et al., 2020) in shortcut learning.

In _open-ended_ tasks --- such as designing word problems or discovering novel proofs --- the goal is not only correctness but also diversity and originality. Often, this requires a far-sighted, creative leap of thought. We argue that this requirement is misaligned with the objective of next-token prediction (NTP). To formulate our intuition, we design a suite of minimal algorithmic tasks loosely based on real-world creative endeavors. Concretely, our tasks require an open-ended _stochastic_ planning step that (a) discovers new connections in a knowledge graph (loosely inspired by word-play, humor or drawing analogies) or (b) constructs new patterns (loosely inspired by constructing word problems, puzzles or mysteries). We then conceptually and empirically argue how NTP leads to myopic shortcut-learning and excessive memorization, limiting its ability to generate novel solutions. In contrast, we find that multi-token approaches, namely teacherless training and diffusion models, can overcome these limitations and comparatively excel on our algorithmic test-bed. Orthogonally, we find that creativity in our tasks is greatly improved by training with a random hash prefix (which we dub as ``_{hash-conditioning_''). Thus our work offers a principled, minimal test-bed for studying open-ended forms of intelligence and also a new angle to take a more serious interest in the paradigm of multi-token prediction.

In-context learning (ICL) typically presents a function through a uniform sample of input-output pairs. Here, we investigate how presenting a compositional subtask curriculum in context may alter the computations that the model learns. We design a compositional algorithmic task based on the modular exponential---a double exponential task composed of two single exponential subtasks---and train transformer models to learn the task in-context. We compare the model when trained (a) using an in-context curriculum consisting of single exponential subtasks and, (b) the model trained directly on the double exponential task without such a curriculum. We show that the model trained with a subtask curriculum can perform zero-shot inference on unseen compositional tasks and is more robust given the same context length. We study how the task is represented across the two training regimes, in particular whether subtask information is represented. We find that the model employs different mechanisms, possibly changing through training, in a way modulated by the data properties of the in-context curriculum.

The correct way to quantify predictive uncertainty in neural networks remains a topic of active discussion. In particular, it is unclear whether the state-of-the art entropy decomposition leads to a meaningful representation of model, or *epistemic*, uncertainty (EU) in the light of a debate that pits *ignorance* against *disagreement* perspectives. We aim to reconcile the conflicting viewpoints by arguing that both are valid but arise from different learning situations. Notably, we show that the presence of *shortcuts* is decisive for EU manifesting as disagreement.

Shortcut learning, where machine learning models exploit spurious correlations in data instead of capturing meaningful features, poses a significant challenge to building robust and generalizable models. This phenomenon is prevalent across various machine learning applications, including vision, natural language processing, and speech recognition, where models may find unintended cues that minimize training loss but fail to capture the underlying structure of the data. Vision classifiers based on Convolutional Neural Networks (CNNs), Multi-Layer Perceptrons (MLPs), and Vision Transformers (ViTs) leverage distinct architectural principles to process spatial and structural information, making them differently susceptible to shortcut learning. In this study, we systematically evaluate these architectures by introducing deliberate shortcuts into the dataset that are positionally correlated with class labels, creating a controlled setup to assess whether models rely on these artificial cues or learn actual distinguishing features. We perform both quantitative evaluation by training on the shortcut-modified dataset and testing them on two different test sets—one containing the same shortcuts and another without them—to determine the extent of reliance on shortcuts. Additionally, qualitative evaluation is performed by using network inversion-based reconstruction techniques to analyze what the models internalize in their weights, aiming to reconstruct the training data as perceived by the classifiers. Further we also evaluate susceptibility to shortcuts learning across different learning rates. Our analysis reveals that CNNs at lower learning rates comparatively tend to be reserved against entirely picking up the shortcut features while ViTs almost entirely ignore the distinctive image features in presence of shortcuts.

Recent studies have revealed various manifestations of position bias in transformer architectures, from the 'lost-in-the-middle' phenomenon to attention sinks, yet a comprehensive theoretical understanding of how attention masks and positional encodings shape these biases remains elusive. This paper introduces a novel graph-theoretic framework to analyze position bias in multi-layer attention. Modeling attention masks as directed graphs, we quantify how tokens interact with contextual information based on their sequential positions. We uncover two key insights: First, causal masking inherently biases attention toward earlier positions, as tokens in deeper layers attend to increasingly more contextualized representations of earlier tokens. Second, we characterize the competing effects of the causal mask and relative positional encodings, such as the decay mask and rotary positional encoding (RoPE): while both mechanisms introduce distance-based decay within individual attention maps, their aggregate effect across multiple attention layers – coupled with the causal mask – leads to a trade-off between the long-term decay effects and the cumulative importance of early sequence positions. Through controlled numerical experiments, we not only validate our theoretical findings but also reproduce position biases observed in real-world LLMs. Our framework offers a principled foundation for understanding positional biases in transformers, shedding light on the complex interplay of attention mechanism components and guiding more informed architectural design.

As training datasets grow larger, we aspire to develop models that generalize well to any diverse test distribution, even if the latter deviates significantly from the training data. Various approaches like domain adaptation, domain generalization, and robust optimization attempt to address the out-of-distribution challenge by posing assumptions about the relation between training and test distribution. Differently, we adopt a more conservative perspective by accounting for the worst-case error across all sufficiently diverse test distributions within a known domain. Our first finding is that training on a uniform distribution over this domain is optimal. We also interrogate practical remedies when uniform samples are unavailable by considering methods for mitigating non-uniformity through finetuning and rebalancing. Our theory aligns with previous observations on the role of entropy and rebalancing for o.o.d. generalization. We also provide new empirical evidence across tasks involving o.o.d. shifts which show applicability of our perspective.

Fine-tuned pretrained attention-based models often struggle with generalisation, leading to poor performance on tasks like out-of-domain transfer, distribution shifts, and few-shot learning. This limitation is prevalent across modalities such as speech, text, graphs, and vision. Nonparametric Variational Information Bottleneck (NVIB) is an attention-based information-theoretic regulariser applicable to pretrained models that has been shown to improve generalisation. However, prior work has applied NVIB only to the text modality and without fine-tuning. We investigate whether NVIB’s ability to remove information from pretrained embeddings helps the model avoid spurious correlations with noisy and superficial features during fine-tuning. We are the first to integrate NVIB regularisation during fine-tuning across multiple diverse models and modalities. This required modifications to the architecture which enhance adaptability and stability during fine-tuning and simplify the evaluation. We found improved out-of-distribution generalisation in: speech quality assessment and language identification, text with induced attention sparsity, graph-based link prediction, and few-shot image classification.

Rich Feature Learning (RFL) aims to extract all beneficial features from the training distribution and has demonstrated significant efficacy in Out-of-Distribution (OOD) generalization. Despite its success, a precise and comprehensive definition of ``richness'' remains elusive. Through an in-depth analysis of RFL algorithms and empirical risk minimization (ERM), the standard OOD baseline, we identify feature diversity as the key differentiator driving RFL's superior OOD performance. Building on this insight, we formally define rich features as those that exhibit both high informativeness and diversity. Leveraging this foundation, we propose Diversity-fOunded Rich fEature lEarniNg (DOREEN), a simple yet highly effective RFL algorithm. We theoretically demonstrate that DOREEN not only realizes the benefits of RFL but also addresses the limitations of prior RFL algorithms. Extensive experiments validate that DOREEN learns richer features and consistently enhances OOD performance across various OOD objectives.

Language models (LMs) deployed in real-world tasks -- such as medical report synthesis, web navigation, and summarization -- must process diverse inputs and handle conflicting information. Users expect them to detect *in-context knowledge conflicts* -- direct contradictions about objective facts -- and issue alerts. Yet, we find a critical failure: when faced with conflicting evidence across **heterogeneous contexts**, such as multiple languages or modalities, LMs fail to detect conflicts, leaving them vulnerable to attacks and misinformation. While they achieve near-perfect accuracy in homogeneous contexts, this **drops by up to 65%** in heterogeneous settings. We identify *context imbalance* as the root cause: LMs exhibit extreme attention asymmetry across domains, disproportionately prioritizing certain domains in mixed inputs. Current instruction-tuning, which trains on separate examples from multiple domains, fails to correct this. To address this, we need *instance-level diverse points* that require reasoning over multiple domains within a single context. We introduce **Heterogeneous Instruction-Tuning (HeteroIT)**, a scalable dataset-mixing procedure that generates instance-level diversity by combining datasets from different domains. Applying HeteroIT to Bactrian-X, a standard multilingual instruction-tuning dataset, improves conflict detection by 37%.

Fine-tuning large-scale pre-trained models often improves in-distribution (ID) performance at the cost of out-of-distribution (OOD) generalization due to overfitting to ID-specific features. To mitigate this, we propose **PCA Dropout**, a novel fine-tuning strategy that suppresses ID-specific feature dependencies by leveraging Principal Component Analysis (PCA). Our method identifies dominant feature components that contribute the most to ID variance and applies structured dropout to reduce their influence, encouraging the model to learn more generalizable representations. We evaluate PCA Dropout on DomainNet and iWildCam using CLIP-based models, demonstrating consistent improvements in OOD robustness over state-of-the-art fine-tuning methods while maintaining strong ID accuracy. Ablation studies further confirm that structured dropout at the feature level outperforms unstructured feature suppression and random dropout strategies.

Generative models are not immune to spurious correlations. The spuriousness in generative models is defined by their ability to compose attributes faithfully, often referred to as compositionality in generative models. To compose attributes successfully, a model should learn to accurately capture the statistical independence between attributes. This paper shows that standard conditional diffusion models violate this assumption, even when all attribute compositions are observed during training. And, this violation is significantly more severe when only a subset of the compositions is observed. We propose CoInD to address this problem. It explicitly enforces statistical independence between the conditional marginal distributions by minimizing Fisher’s divergence between the joint and marginal distributions. The theoretical advantages of CoInD are reflected in both qualitative and quantitative experiments, demonstrating a significantly more faithful and precise controlled generation of samples for arbitrary compositions of attributes

Large Language Models (LLMs) have made significant strides in generating human-like responses, largely due to preference alignment techniques. However, these methods often assume unbiased human feedback, which is rarely the case in real-world scenarios. This paper introduces Bias-Resilient Preference Optimization (BRPO), a novel framework that addresses multiple sources of content-dependent bias in preference learning. BRPO employs a multi-objective optimization approach to separate true preferences from biases, effectively mitigating their impact. We leverage backdoor attack mechanisms to efficiently learn and control for various biases within a single model. Theoretical analysis and extensive experiments on both synthetic and real-world datasets demonstrate that BRPO significantly improves alignment with primary human preferences while controlling for secondary biases such as response length and harmfulness.

Classifiers are used throughout industry to enforce policies, ranging from the detection of toxic content to age-appropriate content filtering. While these classifiers serve important functions, it is also essential that they are built in ways that minimize unfair biases for users. One such fairness consideration is called group fairness, which desires that different sub-population of users receive equal treatment. This is a well-studied problem in the context of classical classifiers. However, the emergence of prompt-based language model (LM) decision making has created new opportunities to solve text-based classification tasks, and the fairness properties of these new classifiers are not yet well understood. Further, the remediation toolkit is incomplete for LM-based decision makers and little is understood about how to improve decision maker group fairness while maintaining classifier performance. This work sets out to add more tools to that toolbox. We introduce adaptations of existing effective approaches from the classical classifier fairness to the prompt-based classifier space. We also devise simple methods that take advantage of the new structure of prompt-based decision makers and operate at the prompt level. We compare these approaches empirically on real data. Our results suggest that adaptations of approaches that are effective for classical classifiers remain effective in the LM-based classifier environment. However, there is room for further exploration of prompt-based remediation methods (and other remediation methods that take advantage of LM structure).

Spurious correlations arise when AI models capture statistical dependencies that do not reflect the true causal structure of the underlying reality, leading to unreliable predictions and unsafe decision-making, particularly in high-stakes domains. While causal discovery methods exist to infer causal structure from data, many are computationally expensive and non-differentiable, limiting their integration into modern AI systems. In this work, we introduce a differentiable approach to causal ordering that allows causal discovery to be seamlessly incorporated as a module within existing machine learning pipelines. Our method builds upon Intersort (Chevalley et al., 2025), a score-based algorithm for discovering causal order in Directed Acyclic Graphs (DAGs) using interventional data. To enable differentiable optimization, we develop a continuous relaxation of Intersort using differentiable sorting and ranking techniques, allowing causal constraints to be directly integrated into gradient-based learning frameworks. By incorporating causal discovery as a regularizer, our approach encourages models to rely on causal relationships rather than spurious correlations, ultimately improving their robustness and trustworthiness when actions are taken based on the learned model. Empirical results demonstrate that enforcing causal order as an inductive bias enhances model generalization and interpretability, making AI systems more reliable and safer for real-world deployment.

Last-layer retraining (LLR) methods — wherein the last layer of a neural network is reinitialized and retrained on a held-out set following ERM training — have recently garnered interest as an efficient approach to rectify dependence on spurious correlations and improve performance on minority groups. Surprisingly, LLR has recently been found to improve worst-group accuracy even when the held-out set is an imbalanced subset of the training set. We initially hypothesize that this “unreasonable effectiveness” of LLR is explained by its ability to mitigate neural collapse through the held-out set, resulting in the implicit bias of gradient descent benefiting robustness. Our empirical investigation does not support this hypothesis. Instead, we present strong evidence for an alternative hypothesis: that the success of LLR is primarily due to better group balance in the held-out set. We conclude by showing how the recent algorithms CB-LLR and AFR perform implicit group-balancing to elicit a robustness improvement.

Action recognition models have shown promising results in understanding consecutive human actions in instructional videos. However, they often rely on dominant action patterns in datasets rather than achieving true video comprehension. We define this as ordinal bias, a systematic reliance on dataset-specific action sequences. To mitigate this, we introduce two simple yet effective video manipulation techniques: action masking and sequence shuffling, where the latter action in dominant pairs is masked, or the sequence is randomized. Our findings reveal that existing models still tend to rely on dominant action pairs and struggle to adapt, highlighting their overestimated performance and lack of robustness.

This work reexamines conventional views of how neural networks store and transfer memorized information by investigating knowledge distillation for random, unstructured data. While knowledge distillation typically focuses on transferring generalizable patterns, we demonstrate that teacher models can encode and transfer purely memorized associations on finite random i.i.d. datasets. Through systematic experiments with fully connected networks, we show that students trained on teacher logits or embedding similarities achieve non-trivial accuracy on memorized data they never directly observed. This phenomenon persists across varying network capacities, dataset compositions, and even with randomized real-world data. Our findings encourage moving beyond simple key-value views of memory in neural networks, and highlight the role of spurious yet learnable patterns that transfer across models—we call them neural mnemonics.

Domain generalization (DG) addresses the challenge of training machine learning models that generalize effectively to unseen target domains exhibiting distributional shifts. Traditional data augmentation techniques, while useful, often fail to adequately simulate the novel domain characteristics necessary for robust DG. We introduce a novel data augmentation framework leveraging the synergistic power of Large Language Models (LLMs) and diffusion models to generate diverse and realistic training data for DG. Our method employs LLMs to create creative prompts that encapsulate new domain styles, which are then used by diffusion models to synthesize high-fidelity images representative of these unseen domains. Furthermore, we integrate a CLIP-guided diversity analysis to ensure that the generated data effectively enhances model generalization while maintaining computational efficiency. Experiments on the PACS dataset show that our method significantly outperforms traditional techniques.

Out-of-Domain (OOD) generalization is the ability of a model trained on one or more domains to generalize to unseen domains. In the ImageNet era of computer vision, evaluation sets for measuring a model's OOD performance were designed to be strictly OOD with respect to style. However, the emergence of foundation models and expansive web-scale datasets has obfuscated this evaluation process, as datasets cover a broad range of domains and risk test domain contamination. In search of the forgotten domain generalization, we create large-scale datasets subsampled from LAION---LAION-Natural and LAION-Rendition---that are strictly OOD to corresponding ImageNet and DomainNet test sets in terms of style. Training CLIP models on these datasets reveals that a significant portion of their performance is explained by in-domain examples. This indicates that the OOD generalization challenges from the ImageNet era still prevail and that training on web-scale data merely creates the illusion of OOD generalization. Furthermore, through a systematic exploration of combining natural and rendition datasets in varying proportions, we identify optimal mixing ratios for model generalization across these domains. Our datasets and results re-enable meaningful assessment of OOD robustness at scale---a crucial prerequisite for improving model robustness.

Iterative decision-making has been widely studied in human cognition and is recognized for its energy efficiency and suitability for biological computations. In contrast, instance segmentation models adopt strategies that diverge from human vision, each presenting unique strengths and limitations. In this paper, we examine the grouping problem in segmentation models and demonstrate that iterative recurrent processing facilitates the identification of diverse solutions and can enhance grouping capabilities. Our experiments further reveal that recurrent processing accelerates convergence and can generate diverse solutions that can help mitigate suboptimal spurious minima. Our work focuses on confounding cases, which have become increasingly relevant as systems are increasingly deployed in safety-critical environments.

With the rapid advancement of text-conditioned Video Generation Models (VGMs), the quality of generated videos has significantly improved, bringing these models closer to functioning as 'world simulators' and making real-world-level video generation more accessible and cost-effective. However, the generated videos often contain factual inaccuracies and lack understanding of fundamental physical laws. While some previous studies have highlighted this issue in limited domains through manual analysis, a comprehensive solution has not yet been established, primarily due to the absence of a generalized, automated approach for modeling and assessing the causal reasoning of these models across diverse scenarios. To address this gap, we propose an automated framework for modeling, evaluating, and measuring the causal understanding of VGMs in real-world scenarios. By combining causal analysis techniques with a carefully designed large language model assistant, our system can assess the causal behavior of models in various contexts without human annotation, which offers strong generalization and scalability. Additionally, we introduce multi-level causal evaluation metrics to provide a detailed analysis of the causal performance of VGMs. As a demonstration, we use our framework to benchmark several prevailing VGMs, offering insight into their causal reasoning capabilities. Our work lays the foundation for systematically addressing the causal understanding deficiencies in VGMs and contributes to advancing their reliability and real-world applicability.

This work concerns the path-star task, a minimal example of searching over a graph. The graph, $G$, is star-shaped with $D$ arms radiating from a start node, $s$. A language model (LM) is given $G$, $s$, and a target node, $t$, which ends one of the arms and is tasked with generating the arm containing $t$. The minimal nature of this task means only a single choice needs to be made: which of the $D$ arms contains $t$? Decoder-only LMs fail to solve this simple task above $1/D$ chance due to a learned shortcut that absorbs training supervision. We show how this pathology is caused by excess supervision and present a series of solutions demonstrating that the task is solvable via decoder-only LMs. We find that the task's minimal nature causes its difficulty, as it prevents task decomposition. Our solutions provide insight into the pathology and its implications for LMs trained via next-token prediction.

As our understanding of the mechanisms of brain function is enhanced, the value of insights gained from neuroscience to the development of AI algorithms deserves further consideration. Here, we draw parallels with an existing tree-based ANN architecture and a recent neuroscience study arguing that the error-based organization of neurons in the cerebellum that share a preference for a personalized view of the entire error space, may account for several desirable features of behavior and learning. We then analyze the learning behavior and characteristics of the model under varying scenarios to gauge the potential benefits of a similar mechanism in ANN. Our empirical results suggest that having separate populations of neurons with personalized error views can enable efficient learning under class imbalance and limited data, and reduce the susceptibility to unintended shortcut strategies, leading to improved generalization. This work highlights the potential of translating the learning machinery of the brain into the design of a new generation of ANNs and provides further credence to the argument that biologically inspired AI may hold the key to overcoming the shortcomings of ANNs.

The problem of spurious correlations (SCs) arises when a classifier relies on non-predictive features that happen to be correlated with the labels in the training data. Previous SC benchmark datasets suffer from varying issues, e.g., over-saturation or only containing one-to-one (O2O) SCs, but no many-to-many (M2M) SCs arising between groups of spurious attributes and classes. In this paper, we present Spawrious-\{O2O, M2M\}-\{Easy, Medium, Hard\}, an image classification benchmark suite containing spurious correlations between classes and backgrounds. We employ a text-to-image model to generate photo-realistic images and an image captioning model to filter out unsuitable ones. The resulting dataset is of high quality and contains approximately 152k images. Our experimental results demonstrate that state-of-the-art group robustness methods struggle with Spawrious.

In settings where both spurious and causal predictors are available, standard neural networks trained under the objective of empirical risk minimization (ERM) with no additional inductive biases tend to have a dependence on a spurious feature. As a result, it is necessary to integrate additional inductive biases in order to guide the network toward generalizable hypotheses. Often these spurious features are shared across related tasks, such as estimating disease prognoses from image scans coming from different hospitals, making the challenge of generalization more difficult. In these settings, it is important that methods are able to integrate the proper inductive biases to generalize across both nuisance-varying families as well as task families. Motivated by this setting, we present RIME (Robustly Informed Meta lEarning), a new method for meta learning under the presence of both positive and negative inductive biases (what to learn and what not to learn). We first develop a theoretical causal framework showing why existing approaches at knowledge integration can lead to worse performance on distributionally robust objectives. We then show that RIME is able to simultaneously integrate both biases, reaching state of the art performance under distributionally robust objectives in informed meta-learning settings under nuisance-varying families.

While multimodal large language models (MLLMs) exhibit remarkable capabilities in visual and textual understanding, they remain highly susceptible to spurious correlations. We propose SpurLens, a novel pipeline leveraging LLMs and open-set object detectors to identify spurious cues and measure their effect on MLLMs in an object detection scenario. Furthermore, we tested different prompting strategies to mitigate this issue, but none proved effective. These findings highlight the urgent need for robust solutions to address spurious correlations in MLLMs.

Evaluating the cognitive abilities of Multi-modal Language Models (MLLMs) is challenging due to their reliance on spurious correlations. To distinguish shortcut-taking from genuine reasoning, we introduce Concept Hacking, a paradigm manipulating concept-relevant information to flip the ground-truth but preserving concept-irrelevant confounds. For instance, in a perceptual constancy test, models must recognize that a uniformly wide bridge does not narrow in the distance; the manipulated condition using concept hacking altered the bridge to actually taper. We assessed 209 models across 45 experiment pairs spanning nine low-level cognitive abilities, encompassing all five core knowledge domains. Comparing performance on manipulated versus standard conditions revealed that models fell into shortcut-reliant or illusory understanding types, with none approaching human-level performance. Models of varying sizes appear in each category, indicating that scaling neither imparts core knowledge nor reduces shortcut reliance. These findings highlight fundamental limitations in current MLLMs, reinforcing concerns about their ability to achieve genuine understanding.

Identifying invariant features – those that stably predict the outcome across diverse environments – is crucial for improving model generalization and uncovering causal mechanisms. While previous methods primarily address this problem through hypothesis testing or regularized optimization, they often lack a principled characterization of the underlying data generative process and struggle with high-dimensional data. In this work, we develop a Bayesian model that encodes an invariance assumption in the generative process of multi-environment data. Within this framework, we perform posterior inference to estimate the invariant features and establish theoretical guarantees on posterior consistency and contraction rates. To address the challenges in high-dimensional settings, we design a scalable variational inference algorithm. We demonstrate the superior inference accuracy and scalability of our method compared to existing approaches in simulations and a gene-perturbation study.

Existing methods for detecting and correcting spurious correlations in image recognition models often fail to identify biasing features due to incoherent groupings of biased images. There is also little exploration of targeted removal of spurious correlations in a low-dimensional feature space. To address these gaps, we propose Performance-Based Feature Sampling (PBFS), a systematic method for producing image recognition models that are de-biased w.r.t a given feature space. We introduce a method for producing coherent bias group proposals (i.e., semantically related images potentially sharing biasing feature(s)) and decorrelating biasing features from the target label using adaptive resampling. We demonstrate that our framework is able to correct for known spurious correlations, and through both established and our proposed metrics, we show that our method is able to de-bias image recognition models both w.r.t a high-dimensional feature space capturing complex representations and w.r.t low-dimensional feature spaces representing simple physical properties.

Classifiers trained with Empirical Risk Minimization (ERM) often rely on spurious correlations, degrading performance on underrepresented groups and challenging out-of-distribution generalization and fairness. While prior methods aim to address this, many require group annotations for training or validation, limiting their applicability when spurious correlations or group labels are unknown. We demonstrate that what has been learned during ERM training can be utilized to \textit{fully} remove group supervision for both training and model selection. To show this, we design Environment-based Validation and Loss-based Sampling (EVaLS), which uses losses from an ERM-trained model to construct datasets with mitigated group imbalance. EVaLS leverages environment inference to create diverse environments with correlation shifts, enabling model selection without group-annotated validation data. By using worst environment accuracy as a tuning surrogate, EVaLS achieves robust performance across groups through simple last-layer retraining. This fast and effective approach eliminates the need for group annotations, achieving competitive worst-group accuracy and improving robustness to known and unknown spurious correlations.

Personalized machine learning models have gained significant importance in various domains, including healthcare. However, designing efficient personalized models remains a challenge. Traditional approaches often involve training multiple sub-models for different population sub-groups, which can be costly and does not always guarantee improved performance across all sub-groups. This paper presents a novel approach to improving model performance at the sub-group level by leveraging bias and training a joint model. Our method involves a two-step process: first, we train a model to predict group attributes, and then we use this model to learn data-dependent biases to modulate a second model for diagnosis prediction. Our results demonstrate that this joint architecture achieves consistent performance gains across all sub-groups in the Heart dataset. Furthermore, in the mortality dataset, it improves performance in two of the four sub-groups. A comparison of our method with the traditional decoupled personalization method demonstrated a greater performance gain in the sub-groups with less harm. This approach offers a more effective and scalable solution for personalization of models, which could have positive impact in healthcare and other areas that require predictive models which take sub-group information into account.

Misinformation is a complex societal issue, and mitigating solutions are difficult to create due to data deficiencies. To address this problem, we have curated the largest collection of (mis)information datasets in the literature, totaling 75. From these, we evaluated the quality of all of the 36 datasets that consist of statements or claims, as well as the 9 datasets that consists of data in purely paragraph form. We assess these datasets to identify those with solid foundations for empirical work and those with flaws that could result in misleading and non-generalizable results, such as insufficient label quality, spurious correlations. We further provide state-of-the-art baselines on all these datasets, but show that regardless of label quality, categorical labels may no longer give an accurate evaluation of detection model performance. We discuss alternatives to mitigate this problem. Overall, this guide aims to provide a roadmap for obtaining higher quality data and conducting more effective evaluations, ultimately improving research in misinformation detection. All datasets and other artifacts are available at [anonymized].

In-context learning (ICL) is the remarkable ability of trained transformers to adapt to new tasks by leveraging a sequence of examples provided at inference time—without any additional training. Prior work on understanding ICL has primarily focused on setups with fixed task complexity (e.g., linear, logistic, or sinusoidal regression tasks with fixed complexity, and more recently first-order Markov chains), overlooking the diverse range of tasks that large language models encounter in practice. In this paper, we investigate ICL in transformers trained on multiple task categories of varying complexity. Our results show that, during inference, transformers effectively learn in-context by identifying the appropriate task complexity and accurately estimating the corresponding task parameters. We verify our claim with experiments on Markov chains and linear regression tasks of varying complexity. Additionally, our experiments suggest that transformers exhibit a bias towards learning the simplest task that explains the inference-time context.

Supervised and preference-based fine-tuning techniques have become popular for aligning large language models (LLMs) with user intent and correctness criteria. However, real-world training data often exhibits spurious correlations—arising from biases, dataset artifacts, or other “shortcut” features—that can compromise a model’s performance or generalization. In this paper, we systematically evaluate three post-training algorithms—Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and KTO (Kahneman-Tversky Optimization)—across a di- verse set of synthetic tasks and spuriousness conditions. Our tasks span mathemat- ical reasoning, constrained instruction-following, and document-grounded ques- tion answering. We vary the degree of spurious correlation (10% vs. 90%) and investigate two forms of artifacts: “Feature Ambiguity” and “Distributional Nar- rowness.” Our results show that the models often but not always degrade under higher spuriousness. The preference-based methods (DPO/KTO) can demonstrate relative robustness in mathematical reasoning tasks. By contrast, SFT maintains stronger performance in complex, context-intensive tasks. These findings high- light that no single post-training strategy universally outperforms in all scenarios; the best choice depends on the type of target task and the nature of spurious cor- relations.

Out-of-distribution (OOD) data often undermines reliable model deployment in high-stakes domains such as financial markets, where overlooked correlations and unexpected shifts can render predictive systems ineffective. We propose STAR (Structured Transformations and Adversarial Reweighting), a framework that leverages the geometry of distribution shifts by combining transformation- based invariances with divergence-based robust optimization. Specifically, STAR places an f -divergence ball around each label-preserving transformation of the training sample, empowering an adversary to apply known transformations and reweight the resulting data within a specified divergence radius. This design cap- tures both large, structured shifts and subtle, unmodeled perturbations—a critical step toward mitigating shortcuts and spurious correlations. Notably, STAR recov- ers standard distributionally robust optimization if no structured transformations are assumed. We establish a uniform-convergence analysis showing that minimiz- ing STAR’s empirical nested min–max objective achieves low worst-case error over all admissible shifts with high probability. Our results quantify the additional samples needed to handle the adversary’s flexibility, providing theoretical guid- ance for selecting the divergence radius based on problem complexity. Empirical studies on synthetic and image benchmarks confirm that STAR outperforms base- lines, consistent with our theoretical findings.

This work addresses the limitations of deep neural networks (DNNs) in generalizing beyond training data due to spurious correlations. Recent research has demonstrated that models trained with empirical risk minimization learn both core and spurious features, often upweighting spurious ones in the final classification, which can frequently lead to poor performance on minority groups. Deep Feature Reweighting alleviates this issue by retraining the model's last classification layer using a group-balanced held-out validation set. However, relying on spurious feature labels during training or validation limits practical application, as spurious features are not always known or costly to annotate. Our preliminary experiments reveal that ERM-trained models exhibit higher gradient norms on minority group samples in the hold-out dataset. Leveraging these insights, we propose an alternative approach called GradTune, which fine-tunes the last classification layer using high-gradient norm samples. Our results on four well-established benchmarks demonstrate that the proposed method can achieve competitive performance compared to existing methods without requiring group labels during training or validation.

This work evaluates Named Entity Recognition (NER) across five large language models (LLMs) using real-world narratives from healthcare and general-purpose datasets, focusing on occupational biases and cross-domain robustness. While prior studies have primarily examined biases in name-based entities using short sentence templates, we shift the focus to evaluating occupational NER in long note templates, analyzing biases across gender, race, and annual wage dimensions. Additionally, we assess cross-domain performance to understand how well the models generalize to unseen domain-specific data, such as healthcare datasets. Our evaluation demonstrates the effectiveness of fine-tuning on domain-specific datasets in improving performance compared to zero-shot and universal NER models. However, significant disparities in model performance and bias representation are observed, highlighting the need for targeted mitigation strategies to ensure subgroup robustness in real-world NER applications.

Object-centric learning (OCL) aims to represent each object's information independently and minimize interference from backgrounds and other objects. OCL is expected to aid model generalization, especially in out-of-distribution (OOD) settings. However, the community's effort has been focused on improving unsupervised entity segmentation performances which is secondary to the main objective. We challenge this. We argue that segmentation is no longer the main barrier: recent class-agnostic segmentation methods reliably localize objects in a zero-shot manner. Instead, we advocate for a renewed emphasis on how decomposed representations can improve OOD generalization. As a first step, we propose Object-Centric Classification with Applied Masks (OCCAM) that exploits discovered objects to extract their representations for downstream classification tasks. Our experiments on datasets with background spurious correlations suggest that even in this task OCL representations do not lead to better generalization than object-centric representations provided by foundational segmentation models. These results showcase the importance of recognizing advances in zero-shot image segmentation when high-performant object-centric representations are the end goal. In addition to that, we suggest exploring new benchmarks for OCL methods evaluation that better reflect the problems these methods are designed to solve and highlight scenarios where OCL methods are more favorable solutions than foundational segmentation models.

This article presents a novel Cross-Instance Contrastive Masking-Enhanced Vision Transformer (CICM-ViT) for hyperspectral image (HSI) classification, which attempts to reduce shortcut learning through Cross-Instance Contrastive Masking (CICM) to enhance spectral-spatial feature extraction through self-supervision. Using the dependencies between instances, CICM-ViT dynamically masks spectral patches across instances, promoting the learning of discriminative features while reducing redundancy, especially in low-data settings. This approach reduces shortcut learning by focusing on global patterns rather than relying on local spurious correlations. CICM-ViT achieves state-of-the-art performance on HSI datasets, with 99.91% OA on Salinas, 96.88% OA on Indian Pines, and 98.88% OA on Botswana, outperforming fourteen SOTA CNN- and transformer-based approaches in both accuracy and efficiency, with only 89,680 parameters.

Deep neural networks trained with Empirical Risk Minimization (ERM) are prone to rely on simple spurious features—features that are correlated with the target but are not causally related to it. To mitigate this over-reliance, Deep Feature Reweighting (DFR) has emerged as an efficient approach, which works by retraining the last layer of an ERM model on a small reweighting dataset. While effective, DFR requires group annotations to create the reweighting dataset, which may be challenging and costly to obtain. Though subsequent works have proposed ways to alleviate this constraint, existing methods still largely rely on group annotations for hyperparameter tuning to achieve robust performance. In this paper, we present LACER, a method that improves group robustness without requiring explicit group annotations for either training or model selection. LACER operates in two stages: first estimating group labels through a loss-weighted clustering formulation that effectively identifies clusters corresponding to underrepresented groups in the validation set, then leveraging these estimated labels for last-layer retraining. Our results provide the empirical evidence that combining semantic feature information with loss values enables effective group label estimation. We validate LACER across multiple vision spurious correlations benchmarks, demonstrating performance comparable to oracle last-layer retraining methods that utilize ground-truth group annotations.

While compositional generalization is fundamental to human intelligence, we still lack understanding of how neural networks combine learned representations of parts into novel wholes. We investigate whether neural networks express representations as linear sums of simpler constituent parts. Our analysis reveals that models trained from scratch often exhibit decodability, where the features can be linearly decoded to perform well, but may lack linear structure, preventing the models from generalizing zero-shot. Instead, linearity of representations only arises with high training data diversity. We prove that when representations are linear, perfect generalization to novel concept combinations is possible with minimal training data. Empirically evaluating large-scale pretrained models through this lens reveals that they achieve strong generalization for certain concept types while still falling short of the ideal linear structure for others.

Compositional generalization—the ability to reason about novel combinations of familiar concepts—is fundamental to human cognition and a critical challenge for machine learning. Object-Centric representation learning has been proposed as a promising approach for achieving this capability. However, systematic evaluation of these methods in visually complex settings remains limited. In this work, we introduce a benchmark to measure how well vision encoders, with and without object-centric biases, generalize to unseen combinations of object properties. Using CLEVRTex-style images, we create multiple training splits with partial coverage of object property combinations and generate question--answer pairs to assess compositional generalization on a held-out test set. We focus on comparing pretrained foundation models with object-centric models that incorporate such foundation models as backbones---a leading approach in this domain. To ensure a fair and comprehensive comparison, we carefully account for representation format differences. In this preliminary study, we use DINOv2 as the foundation model and DINOSAURv2 as its object-centric counterpart. We control for compute budget and differences in image representation sizes to ensure robustness. Our key findings reveal that object-centric approaches (1) converge faster on in-distribution data but underperform slightly when non-object-centric models are given a significant compute advantage, and (2) they exhibit superior compositional generalization, outperforming DINOv2 on unseen combinations of object properties while requiring approximately four to eight times less downstream compute.

Neural networks (NNs) are known to exhibit simplicity bias, where they tend to prioritize learning simple features over more complex ones, even when the latter are more informative. This bias can result in models making skewed predictions with poor out-of-distribution (OOD) generalization. To address this issue, we propose three techniques to mitigate simplicity bias. One of these is a modification to the Feature Sieve method. In the second method we utilize neuronal correlations as a penalizing effect to try and enforce the learning of different features. The third technique involves a novel feature-building approach called Self-Supervised Augmentation. We validate our methods' generalization capabilities through experiments on a custom dataset.

The performance of Large Vision-Language Models (LVLMs) in In-Context Learning (ICL) is heavily influenced by the quality of ICL sequences, particularly in tasks requiring cross-modal reasoning and open-ended generation. To address this challenge, we innovatively interpret multimodal ICL from the perspective of task mapping. We systematically model local and global relationships within in-context demonstrations (ICDs) and demonstrate their core role and cohesion in enhancing LVLM performance. Inspired by these findings, we propose Ta-ICL, a lightweight transformer-based model equipped with task-aware attention to dynamically configure ICL sequences. By integrating task mapping into the autoregressive process, Ta-ICL achieves bidirectional enhancement between sequence configuration and task reasoning. Through extensive experiments, we demonstrate that Ta-ICL effectively improves multimodal ICL across various LVLMs and tasks. Our results highlight the potential of task mapping to be widely applied in enhancing multimodal reasoning, paving the way for robust and generalizable multimodal ICL frameworks.