Machine learning applications cover a wide range of predictive tasks in which tabular datasets play a significant role. However, although they often address similar problems, tabular datasets are typically treated as standalone tasks. The possibilities of using previously solved problems are limited due to the lack of structured contextual information about their features and the lack of understanding of the relations between them. To overcome this limitation, we propose a new approach called Semantic Feature Net (SeFNet), capturing the semantic meaning of the analyzed tabular features. By leveraging existing ontologies and domain knowledge, SeFNet opens up new opportunities for sharing insights between diverse predictive tasks. One such opportunity is the Dataset Ontology-based Semantic Similarity (DOSS) measure, which quantifies the similarity between datasets using relations across their features. In this paper, we present an example of SeFNet prepared for a collection of predictive tasks in healthcare, with the features' relations derived from the SNOMED-CT ontology. The proposed SeFNet framework and the accompanying DOSS measure address the issue of limited contextual information in tabular datasets. By incorporating domain knowledge and establishing semantic relations between features, we enhance the potential for meta-learning and enable valuable insights to be shared across different predictive tasks.

Predictive models may generate biased predictions when classifying imbalanced datasets. This happens when the model favors the majority class, leading to low performance in accurately predicting the minority class. To address this issue, balancing or resampling methods are critical data-centric AI approaches in the modeling process to improve prediction performance. However, there have been debates and questions about the functionality of these methods in recent years. In particular, many candidate models may exhibit very similar predictive performance, called the Rashomon effect, in model selection, and they may even produce different predictions for the same observations. Selecting one of these models without considering the predictive multiplicity -- which is the case of yielding conflicting models' predictions for any sample -- can result in blind selection. In this paper, the impact of balancing methods on predictive multiplicity is examined using the Rashomon effect. It is crucial because the blind model selection in data-centric AI is risky from a set of approximately equally accurate models. This may lead to severe problems in model selection, validation, and explanation. To tackle this matter, we conducted real dataset experiments to observe the impact of balancing methods on predictive multiplicity through the Rashomon effect by using a newly proposed metric obscurity in addition to the existing ones: ambiguity and discrepancy. Our findings showed that balancing methods inflate the predictive multiplicity and yield varying results. To monitor the trade-off between the prediction performance and predictive multiplicity for conducting the modeling process responsibly, we proposed using the extended version of the performance-gain plot when balancing the training data.

This paper introduces a new modeling perspective in the public mental health domain to provide a robust interpretation of the relations between anxiety and depression, and the demographic and temporal factors. This perspective particularly leverages the Rashomon Effect, where multiple models exhibit similar predictive performance but rely on diverse internal structures. Instead of considering these multiple models, choosing a single best model risks masking alternative narratives embedded in the data. To address this, we employed this perspective in the interpretation of a large-scale psychological dataset, specifically focusing on the Patient Health Questionnaire-4. We use a random forest model combined with partial dependence profiles to rigorously assess the robustness and stability of predictive relationships across the resulting Rashomon set, which consists of multiple models that exhibit similar predictive performance. Our findings confirm that demographic variables , , and lead to consistent structural shifts in anxiety and depression risk. Crucially, we identify significant temporal effects: risk probability demonstrates clear diurnal and circaseptan fluctuations, peaking during early morning hours. This work demonstrates the necessity of moving beyond the best model to analyze the entire Rashomon set. Our results highlight that the observed variability, particularly due to circadian and circaseptan rhythms, must be meticulously considered for robust interpretation in psychological screening. We advocate for a multiplicity-aware approach to enhance the stability and generalizability of ML-based conclusions in mental health research.

Vision foundation models, such as DINOv2, learn highly expressive representations but rely on massive, opaque architectures that demand substantial computational power and memory. To provide an interpretable-guided and efficient solution to this issue, we first propose a spectral analysis and new visualization technique for individual attention heads based on the Laplacian eigenvectors of their attention maps. Building upon recent observations regarding the block structure of Vision Transformers, we perform semantic clustering of attention heads and identify functional redundancies. Leveraging these insights, we introduce SAPER (Soft Attention PrunER), an end-to-end differentiable pruning framework based on the LapSum Soft Top-K approach. Extensive experiments on ImageNet-1K demonstrate that SAPER achieves a highly favorable accuracy-efficiency trade-off, outperforming the competitive RAPTOR baseline in FLOPs reduction while preserving strong classification performance.

Interpreting individual neurons in deep neural networks is a crucial step towards understanding their complex decision-making processes and ensuring AI safety. Despite recent progress in neuron labeling, existing methods often limit the search space to predefined concept vocabularies or produce overly specific descriptions that fail to capture higher-order, global concepts. We introduce LINE, a novel, training-free iterative approach tailored for open-vocabulary concept labeling in vision models. Operating in a strictly black-box setting, LINE leverages a large language model and a text-to-image generator to iteratively propose and refine concepts in a closed loop, guided by activation history. LINE achieves state-of-the-art performance across multiple model architectures, yielding AUC improvements of up to 0.11 on ImageNet and 0.05 on Places365, while discovering, on average, 27

We introduce the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. For any first-order attribution $ϕ(f)$ explaining a model $f$, we measure the directional influence of feature $j$ on the attribution of feature $i$, denoted as meta-attribution $φ_{j \to i}(f)$, by treating the attribution method itself as a cooperative game and computing its Shapley value. Theoretically, we prove that attributions hierarchically decompose into meta-attributions, and establish these as directional extensions of existing interaction indices. Empirically, we demonstrate that the metagame delivers insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.

Diffusion models are prone to generating structural hallucinations - samples that match the statistical properties of the training data yet defy underlying structural rules, resulting in anomalies like hands with more than five fingers. Recent research studied this failure mode from several viewpoints, offering partial explanations to their occurrence, such as mode interpolation. In this work, we propose a complementary perspective that treats hallucinations as instabilities on the model-induced manifold. We begin by showing that a hallucination filter based on such instabilities matches or exceeds the performance of the recently proposed temporal one. By tracing the source of these instabilities, we identify local intrinsic dimension (LID) as their primary driver and propose Intrinsic Quenching (IQ), a direct corrective mechanism that deflates it to alleviate hallucinations. IQ consistently outperforms standard hallucination reduction baselines across a wide array of benchmarks and offers a highly promising solution for enforcing anatomical consistency in downstream medical imaging tasks.

We argue that the AI community is now ready to move beyond benchmarking and consolidate scattered efforts in model analysis into a systematic discipline, a direction we term Model Science. Complex AI models now serve billions of users, yet our understanding of how they work lags far behind our ability to deploy them. Decades of benchmark-driven research have delivered remarkable progress: extensive leaderboards, a wide range of performance metrics, tracking capability gains across diverse tasks; yet this success has also revealed the limits of benchmarks as they tell us whether models perform but not why they succeed or fail, they miss critical failure modes, such as hallucinations or shortcuts. Precedents from established sciences point the way forward: cognitive science shows that understanding complex systems requires complementary levels of analysis; neuroscience demonstrates that deep study of single cases reveals what population studies miss; medicine teaches that specialised training must develop alongside research practice; and agriculture models how shared infrastructure and principles enable cumulative progress. These lessons inform three foundations for Model Science. First, we propose to consolidate research around four functional perspectives: Verify, Explore, Steer, and Refine that address complementary questions about model behaviour. Second, we discuss the required infrastructure for cumulative knowledge: catalogues of datasets, models and findings. Third, we highlight the need for deep analysis of individual model instances, not just model families, because single cases can reveal what population studies miss.

Forecasting plays a crucial role in modern safety-critical applications, such as space operations. However, the increasing use of deep forecasting models introduces a new security risk of trojan horse attacks, carried out by hiding a backdoor in the training data or directly in the model weights. Once implanted, the backdoor is activated by a specific trigger pattern at test time, causing the model to produce manipulated predictions. We focus on this issue in our Trojan Horse Hunt data science competition, where more than 200 teams faced the task of identifying triggers hidden in deep forecasting models for spacecraft telemetry. We describe the novel task formulation, benchmark set, evaluation protocol, and best solutions from the competition. We further summarize key insights and research directions for effective identification of triggers in time series forecasting models. All materials are publicly available on the official competition webpage https://www.kaggle.com/competitions/trojan-horse-hunt-in-space.

Abstract Transarterial chemoembolization (TACE) is a standard treatment for patients with unresectable hepatocellular carcinoma (HCC), yet existing models provide limited individualized risk stratification. Automated CT-derived body composition analysis has emerged as an objective marker of patient physiological reserve, but its value in prognostication in TACE patients is insufficiently studied. Therefore, the aim of the study was to evaluate the prognostic value of a fully automated, open-source pipeline for CT-based body composition analysis in predicting overall survival (OS) in patients with HCC undergoing TACE. In this study, we used two independent cohorts of treatment-naive patients undergoing TACE: the WAW-TACE cohort (development; n = 230, OS: 28.6 months) and the HCC-TACE-Seg cohort (validation; n = 100, OS: 24.0 months). Skeletal muscle and fat metrics were extracted from pre-treatment CTs using a standardized deep learning pipeline and normalized by sex. Survival analyses were performed using Cox proportional hazards (CoxPH) models and random survival forests (RSF). Skeletal muscle density (SMD) at the L3 level was the strongest and independent predictor of OS across both cohorts (HR: development, 0.84; p = 0.029; validation, 0.79; p = 0.028). This association remained significant after adjustment for the best-performing clinical composite scores: mHAP-2 in the development (adjusted HR = 0.68; p = 0.049) and CLIP in the validation cohort (adjusted HR = 0.43; p = 0.003). In CoxPH, the addition of SMD metrics resulted in only modest improvements in discrimination (ΔC-index 0.011–0.037) that did not reach statistical significance. In contrast, RSF analysis demonstrated a statistically significant improvement in model discrimination when muscle-based variables were added to clinical features (ΔC-index = 0.023; p < 0.001). In both cohorts, SMD showed a reproducible independent prognostic association with overall survival. While adding SMD to traditional clinical models resulted in only modest, and in Cox-based analyses not statistically significant, improvements in discrimination, SMD provided complementary prognostic information. This suggests that the primary value of these automated CT-derived body composition metrics lies not in their performance as standalone predictors, but in their ability to provide an additional layer of objective biological data that may contribute to risk stratification in a complementary and exploratory manner within multivariable frameworks. Notably, in internally cross-validated RSF analyses, statistically significant increases in model discrimination were observed when muscle-based features were integrated into the model, highlighting their potential complementary value within machine learning frameworks.

Contrastively pre-trained Vision-Language Models (VLMs) serve as powerful feature extractors. Yet, their shared latent spaces are prone to structural anomalies and act as repositories for non-semantic, multi-modal noise. To address this phenomenon, we employ spectral decomposition of covariance matrices to decompose the VLM latent space into a multi-modal semantic signal component and a shared noise subspace. We observe that this noise geometry exhibits strong subgroup invariance across distinct data subsets. Crucially, pruning these shared noise dimensions is mainly harmless, preserving or actively improving downstream task performance. By isolating true semantic signals from artifactual noise, this work provides new mechanistic insights into the representational structure of modern VLMs, suggesting that a substantial fraction of their latent geometry is governed by shared, architecture-level noise rather than task-relevant semantics alone.

Scientific knowledge about AI models is produced faster than the community can organize it. Every few months a new foundation model reshapes the field and hundreds of papers, blogs, and technical reports document how each behaves or fails. Yet, these findings remain scattered and effectively unretrievable. To address this gap we present Modelpedia, an automated, LLM-assisted framework that extracts findings about models from published papers, links it to the model, dataset, method, and concept it concerns, and aggregates the result into a searchable public catalog. Applying the prototype to accepted ICLR 2024 and 2025 papers, we extract over a thousand findings and, treating the catalog itself as an object of study, run a meta-analysis of how the community investigates models. Now, we invite the community to explore, contribute to, and build on the open catalog, and to help establish model findings as a shared foundation for the meta-science of AI.

Lung cancer remains the leading cause of cancer mortality, driving the development of automated screening tools to alleviate radiologist workload. Standing at the frontier of this effort is Sybil, a deep learning model capable of predicting future risk solely from computed tomography (CT) with high precision. However, despite extensive clinical validation, current assessments rely purely on observational metrics. This correlation-based approach overlooks the model's actual reasoning mechanism, necessitating a shift to causal verification to ensure robust decision-making before clinical deployment. We propose S(H)NAP, a model-agnostic auditing framework that constructs generative interventional attributions validated by expert radiologists. By leveraging realistic 3D diffusion bridge modeling to systematically modify anatomical features, our approach isolates object-specific causal contributions to the risk score. Providing the first interventional audit of Sybil, we demonstrate that while the model often exhibits behavior akin to an expert radiologist, differentiating malignant pulmonary nodules from benign ones, it suffers from critical failure modes, including dangerous sensitivity to clinically unjustified artifacts and a distinct radial bias.

Explainability is essential for deploying deep learning models in high-stakes medical applications. Existing explainability methods for volumetric imaging predominantly operate in voxel space, overlooking the structured representations introduced by recent advances in 3D scene modeling. We present COGENT (Counterfactual Gaussian Explanations), a framework that generates counterfactual explanations directly in the parameter space of Gaussian-based volumetric representations. Built upon MedGS and the Sybil lung cancer risk prediction model, COGENT optimizes selected Gaussian primitives through a differentiable rendering pipeline, enabling gradients from the downstream predictor to identify representation components that most influence model decisions. Unlike conventional pixel- or voxel-level attribution methods, our approach formulates explainability as a counterfactual optimization problem over an explicit 3D scene representation, producing sparse and spatially localized explanations while preserving anatomical consistency. We evaluate COGENT on lung CT scans using quantitative comparisons with existing explainability methods together with qualitative analysis by medical experts. The results demonstrate that representation-space counterfactual optimization provides clinically meaningful explanations while offering a new perspective on interpreting volumetric deep learning models.

Vision-Language Models (VLMs) are demonstrating significant capabilities in medical tasks like radiology analysis, yet providing faithful and interpretable explanations remains a key consideration for their responsible deployment in clinical settings. However, existing explanation methods, such as the widely used FIxLIP framework, often struggle with the fine-grained nature of modern tokenizers. The tokenization problem fragments clinical concepts---splitting terms like "saddle embolus" into scattered, meaningless subwords---which leads to noisy, semantically incoherent cross-modal attributions. Such fragmentation also results in a combinatorial explosion of interaction possibilities, obscuring the model's true reasoning. To address this, we introduce ParseFIxLIP, an extension that incorporates the Tree-Gram Parsing into the Banzhaf interaction game used by FIxLIP. This semantically informed strategy utilizes dependency parsing trees to define explanation players by grouping related text tokens into semantically coherent units. Our smart_depth grouping strategy, merging tokens according to spaCy token dependency tree, successfully mitigates concept fragmentation, yielding substantially more interpretable cross-modal interactions by unifying complex medical concepts. Quantitatively, while baselines struggled with the high dimensionality of long captions, our parsing approach maintained statistical robustness and semantic parsimony. Qualitative analysis on BiomedCLIP, validated on medical imagery (ROCOv2) and general examples, confirms that the approach accurately captures the synergistic influence of grouped words on model predictions. In conclusion, our work offers intuitive and clinically relevant insights into VLM decision-making, fulfilling the critical need for coherent explanations in the medical domain.

This paper investigates the critical role of hyperparameters in predictive multiplicity, where different machine learning models trained on the same dataset yield divergent predictions for identical inputs. These inconsistencies can seriously impact high-stakes decisions such as credit assessments, hiring, and medical diagnoses. Focusing on six widely used models for tabular data - Elastic Net, Decision Tree, k-Nearest Neighbor, Support Vector Machine, Random Forests, and Extreme Gradient Boosting - we explore how hyperparameter tuning influences predictive multiplicity, as expressed by the distribution of prediction discrepancies across benchmark datasets. Key hyperparameters such as lambda in Elastic Net, gamma in Support Vector Machines, and alpha in Extreme Gradient Boosting play a crucial role in shaping predictive multiplicity, often compromising the stability of predictions within specific algorithms. Our experiments on 21 benchmark datasets reveal that tuning these hyperparameters leads to notable performance improvements but also increases prediction discrepancies, with Extreme Gradient Boosting exhibiting the highest discrepancy and substantial prediction instability. This highlights the trade-off between performance optimization and prediction consistency, raising concerns about the risk of arbitrary predictions. These findings provide insight into how hyperparameter optimization leads to predictive multiplicity. While predictive multiplicity allows prioritizing domain-specific objectives such as fairness and reduces reliance on a single model, it also complicates decision-making, potentially leading to arbitrary or unjustified outcomes.

We discover a theoretical connection between explanation estimation and distribution compression that significantly improves the approximation of feature attributions, importance, and effects. While the exact computation of various machine learning explanations requires numerous model inferences and becomes impractical, the computational cost of approximation increases with an ever-increasing size of data and model parameters. We show that the standard i.i.d. sampling used in a broad spectrum of algorithms for post-hoc explanation leads to an approximation error worthy of improvement. To this end, we introduce Compress Then Explain (CTE), a new paradigm of sample-efficient explainability. It relies on distribution compression through kernel thinning to obtain a data sample that best approximates its marginal distribution. CTE significantly improves the accuracy and stability of explanation estimation with negligible computational overhead. It often achieves an on-par explanation approximation error 2-3x faster by using fewer samples, i.e. requiring 2-3x fewer model evaluations. CTE is a simple, yet powerful, plug-in for any explanation method that now relies on i.i.d. sampling.

Analysis of 3D segmentation models, especially in the context of medical imaging, is often limited to segmentation performance metrics that overlook the crucial aspect of explainability and bias. Currently, effectively explaining these models with saliency maps is challenging due to the high dimensions of input images multiplied by the ever-growing number of segmented class labels. To this end, we introduce Agg^2Exp, a methodology for aggregating fine-grained voxel attributions of the segmentation model's predictions. Unlike classical explanation methods that primarily focus on the local feature attribution, Agg^2Exp enables a more comprehensive global view on the importance of predicted segments in 3D images. Our benchmarking experiments show that gradient-based voxel attributions are more faithful to the model's predictions than perturbation-based explanations. As a concrete use-case, we apply Agg^2Exp to discover knowledge acquired by the Swin UNEt TRansformer model trained on the TotalSegmentator v2 dataset for segmenting anatomical structures in computed tomography medical images. Agg^2Exp facilitates the explanatory analysis of large segmentation models beyond their predictive performance. The source code is publicly available at https://github.com/mi2datalab/agg2exp.

We present a process of transforming medical data into a system that supports radiologists' interpretation and understanding of Computed Tomography (CT) images. The system is based on a pipeline that includes image conversion, organ segmentation, feature extraction, and report rendering. The final report presents organ visualisations and information about organ measurements, with marked outliers, to the radiologist. The system was created using data from the database containing over 40,000 CT scans and a pre-trained Swin UNETR architecture. The system obtained 89.09\% DICE for five segmented organs. The created solution can go through the process in less than five and a half minutes, and its usability was confirmed by radiologists.

Sparse autoencoders (SAEs) are useful for detecting and steering interpretable features in neural networks, with particular potential for understanding complex multimodal representations. Given their ability to uncover interpretable features, SAEs are particularly valuable for analyzing vision-language models (e.g., CLIP and SigLIP), which are fundamental building blocks in modern large-scale systems yet remain challenging to interpret and control. However, current SAE methods are limited by optimizing both reconstruction quality and sparsity simultaneously, as they rely on either activation suppression or rigid sparsity constraints. To this end, we introduce Matryoshka SAE (MSAE), a new architecture that learns hierarchical representations at multiple granularities simultaneously, enabling a direct optimization of both metrics without compromise. MSAE establishes a state-of-the-art Pareto frontier between reconstruction quality and sparsity for CLIP, achieving 0.99 cosine similarity and less than 0.1 fraction of variance unexplained while maintaining 80\% sparsity. Finally, we demonstrate the utility of MSAE as a tool for interpreting and controlling CLIP by extracting over 120 semantic concepts from its representation to perform concept-based similarity search and bias analysis in downstream tasks like CelebA. We make the codebase available at https://github.com/WolodjaZ/MSAE.