arXiv subjects

Wojciech Samek

Publications and source records attributed to Wojciech Samek.

At least 55 records · Page 3Linked to original sources

arXiv2026

Atlas-Alignment: Making Interpretability Transferable Across Language Models

Interpretability is crucial for building safe, reliable, and controllable language models, yet existing interpretability pipelines remain costly and difficult to scale. Interpreting a new model typically requires training model-specific components (e.g., sparse autoencoders), followed by manual or semi-automated labeling and validation, imposing a growing "transparency tax" that does not scale with the pace of model development. We introduce Atlas-Alignment, a framework that avoids this cost by aligning the latent space of a new model to a pre-existing, labeled Concept Atlas using only shared inputs and lightweight representational alignment methods. Through quantitative and qualitative evaluations, we show that simple alignment methods enable robust semantic retrieval and steerable generation without the need for labeled concept datasets. Atlas-Alignment thus amortizes the cost of explainable AI and mechanistic interpretability: by investing in a single high-quality Concept Atlas, we can make many new models transparent and controllable at minimal marginal cost.

cs.LG↗

arXiv2026

From Weights to Activations: Is Steering the Next Frontier of Adaptation?

Post-training adaptation of language models is commonly achieved through parameter updates or input-based methods such as fine-tuning, parameter-efficient adaptation, and prompting. In parallel, a growing body of work modifies internal activations at inference time to influence model behavior, an approach known as steering. Despite increasing use, steering is rarely analyzed within the same conceptual framework as established adaptation methods. In this work, we argue that steering should be regarded as a form of model adaptation. We introduce a set of functional criteria for adaptation methods and use them to compare steering approaches with classical alternatives. This analysis positions steering as a distinct adaptation paradigm based on targeted interventions in activation space, enabling local and reversible behavioral change without parameter updates. The resulting framing clarifies how steering relates to existing methods, motivating a unified taxonomy for model adaptation.

cs.CL↗

arXiv2026

X-SYS: A Reference Architecture for Interactive Explanation Systems

The explainable AI (XAI) research community has proposed numerous technical methods, yet deploying explainability as systems remains challenging: Interactive explanation systems require both suitable algorithms and system capabilities that maintain explanation usability across repeated queries, evolving models and data, and governance constraints. We argue that operationalizing XAI requires treating explainability as an information systems problem where user interaction demands induce specific system requirements. We introduce X-SYS, a reference architecture for interactive explanation systems, that guides (X)AI researchers, developers and practitioners in connecting interactive explanation user interfaces (XUI) with system capabilities. X-SYS organizes around four quality attributes named STAR (scalability, traceability, responsiveness, and adaptability), and specifies a five-component decomposition (XUI Services, Explanation Services, Model Services, Data Services, Orchestration and Governance). It maps interaction patterns to system capabilities to decouple user interface evolution from backend computation. We implement X-SYS through SemanticLens, a system for semantic search and activation steering in vision-language models. SemanticLens demonstrates how contract-based service boundaries enable independent evolution, offline/online separation ensures responsiveness, and persistent state management supports traceability. Together, this work provides a reusable blueprint and concrete instantiation for interactive explanation systems supporting end-to-end design under operational constraints.

cs.AI↗

arXiv2026

Structural Compactness as a Complementary Criterion for Explanation Quality

In the evaluation of attribution quality, the quantitative assessment of explanation legibility is particularly difficult, as it is influenced by varying shapes and internal organization of attributions not captured by simple statistics. To address this issue, we introduce Minimum Spanning Tree Compactness (MST-C), a graph-based structural metric that captures higher-order geometric properties of attributions, such as spread and cohesion. These components are combined into a single score that evaluates compactness, favoring attributions with salient points spread across a small area and spatially organized into few but cohesive clusters. We show that MST-C reliably distinguishes between explanation methods, exposes fundamental structural differences between models, and provides a robust, self-contained diagnostic for explanation compactness that complements existing notions of attribution complexity.

cs.AI↗

arXiv2026

Democratizing Federated Learning with Blockchain and Multi-Task Peer Prediction

The synergy between Federated Learning and blockchain has been considered promising; however, the computationally intensive nature of contribution measurement conflicts with the strict computation and storage limits of blockchain systems. We propose a novel concept to decentralize the AI training process using blockchain technology and Multi-task Peer Prediction. By leveraging smart contracts and cryptocurrencies to incentivize contributions to the training process, we aim to harness the mutual benefits of AI and blockchain. We discuss the advantages and limitations of our design.

cs.CR↗

arXiv2026

Attribution-Guided Decoding

The capacity of Large Language Models (LLMs) to follow complex instructions and generate factually accurate text is critical for their real-world application. However, standard decoding methods often fail to robustly satisfy these requirements, while existing control techniques frequently degrade general output quality. In this work, we introduce Attribution-Guided Decoding (AGD), an interpretability-based decoding strategy. Instead of directly manipulating model activations, AGD considers a set of high-probability output token candidates and selects the one that exhibits the highest attribution to a user-defined Region of Interest (ROI). This ROI can be flexibly defined over different parts of the model's input or internal components, allowing AGD to steer generation towards various desirable behaviors. We demonstrate AGD's efficacy across three challenging domains. For instruction following, we show that AGD significantly boosts adherence (e.g., improving the overall success rate on Llama 3.1 from 66.0% to 79.1%). For knowledge-intensive tasks, we show that guiding generation towards usage of internal knowledge components or contextual sources can reduce hallucinations and improve factual accuracy in both closed-book and open-book settings. Furthermore, we propose an adaptive, entropy-based variant of AGD that mitigates quality degradation and reduces computational overhead by applying guidance only when the model is uncertain. Our work presents a versatile, more interpretable, and effective method for enhancing the reliability of modern LLMs.

cs.LG↗

arXiv2026

Building Trust in PINNs: Error Estimation through Finite Difference Methods

Physics-informed neural networks (PINNs) constitute a flexible deep learning approach for solving partial differential equations (PDEs), which model phenomena ranging from heat conduction to quantum mechanical systems. Despite their flexibility, PINNs offer limited insight into how their predictions deviate from the true solution, hindering trust in their prediction quality. We propose a lightweight post-hoc method that addresses this gap by producing pointwise error estimates for PINN predictions, which offer a natural form of explanation for such models, identifying not just whether a prediction is wrong, but where and by how much. For linear partial differential equations, the error between a PINN approximation and the true solution satisfies the same differential operator as the original problem, but driven by the PINN's PDE residual as its source term. We solve this error equation numerically using finite difference methods requiring no knowledge of the true solution. Evaluated on several benchmark PDEs, our method yields accurate error maps at low computational cost, enabling targeted and interpretable validation of PINNs.

cs.LG↗

arXiv2026

Circuit Insights: Towards Interpretability Beyond Activations

The fields of explainable AI and mechanistic interpretability aim to uncover the internal structure of neural networks, with circuit discovery as a central tool for understanding model computations. Existing approaches, however, rely on manual inspection and remain limited to toy tasks. Automated interpretability offers scalability by analyzing isolated features and their activations, but it often misses interactions between features and depends strongly on external LLMs and dataset quality. Transcoders have recently made it possible to separate feature attributions into input-dependent and input-invariant components, providing a foundation for more systematic circuit analysis. Building on this, we propose WeightLens and CircuitLens, two complementary methods that go beyond activation-based analysis. WeightLens interprets features directly from their learned weights, removing the need for explainer models or datasets while matching or exceeding the performance of existing methods on context-independent features. CircuitLens captures how feature activations arise from interactions between components, revealing circuit-level dynamics that activation-only approaches cannot identify. Together, these methods increase interpretability robustness and enhance scalable mechanistic analysis of circuits while maintaining efficiency and quality.

cs.LG↗

arXiv2026

Dyslexify: A Mechanistic Defense Against Typographic Attacks in CLIP

Typographic attacks exploit multi-modal systems by injecting text into images, leading to targeted misclassifications, malicious content generation and even Vision-Language Model jailbreaks. In this work, we analyze how CLIP vision encoders behave under typographic attacks, locating specialized attention heads in the latter half of the model's layers that causally extract and transmit typographic information to the cls token. Building on these insights, we introduce Dyslexify - a method to defend CLIP models against typographic attacks by selectively ablating a typographic circuit, consisting of attention heads. Without requiring finetuning, dyslexify improves performance by up to 22.06% on a typographic variant of ImageNet-100, while reducing standard ImageNet-100 accuracy by less than 1%, and demonstrate its utility in a medical foundation model for skin lesion diagnosis. Notably, our training-free approach remains competitive with current state-of-the-art typographic defenses that rely on finetuning. To this end, we release a family of dyslexic CLIP models which are significantly more robust against typographic attacks. These models serve as suitable drop-in replacements for a broad range of safety-critical applications, where the risks of text-based manipulation outweigh the utility of text recognition.

cs.CV↗

arXiv2026

ASIDE: Architectural Separation of Instructions and Data in Language Models

Despite their remarkable performance, large language models lack elementary safety features, making them susceptible to numerous malicious attacks. In particular, previous work has identified the absence of an intrinsic separation between instructions and data as the root cause of the success of prompt injection attacks. In this work, we propose a new architectural element, ASIDE, that allows language models to clearly separate instructions and data at the level of token embeddings. ASIDE applies an orthogonal rotation to the embeddings of data tokens, thus creating clearly distinct representations of instructions and data tokens without introducing any additional parameters. As we demonstrate experimentally across a range of models, instruction-tuning LLMs with ASIDE (1) achieves substantially higher instruction-data separation without performance loss and (2) makes the models more robust to prompt injection benchmarks, even without dedicated safety training. Additionally, we provide insights into the mechanism underlying our method through an analysis of the model representations. The source code and training scripts are openly accessible at https://github.com/egozverev/aside.

cs.LG↗

arXiv2026

Towards Visually Explaining Statistical Tests with Applications in Biomedical Imaging

Deep neural two-sample tests have recently shown strong power for detecting distributional differences between groups, yet their black-box nature limits interpretability and practical adoption in biomedical analysis. Moreover, most existing post-hoc explainability methods rely on class labels, making them unsuitable for label-free statistical testing settings. We propose an explainable deep statistical testing framework that augments deep two-sample tests with sample-level and feature-level explanations, revealing which individual samples and which input features drive statistically significant group differences. Our method highlights which image regions and which individual samples contribute most to the detected group difference, providing spatial and instance-wise insight into the test's decision. Applied to biomedical imaging data, the proposed framework identifies influential samples and highlights anatomically meaningful regions associated with disease-related variation. This work bridges statistical inference and explainable AI, enabling interpretable, label-free population analysis in medical imaging.

cs.CV↗

arXiv2026

Concept activation vectors: a unifying view and adversarial attacks

Concept Activation Vectors (CAVs) are a tool from explainable AI, offering a promising approach for understanding how human-understandable concepts are encoded in a model's latent spaces. They are computed from hidden-layer activations of inputs belonging either to a concept class or to non-concept examples. Adopting a probabilistic perspective, the distribution of the (non-)concept inputs induces a distribution over the CAV, making it a random vector in the latent space. This enables us to derive mean and covariance for different types of CAVs, leading to a unified theoretical view. This probabilistic perspective also reveals a potential vulnerability: CAVs can strongly depend on the rather arbitrary non-concept distribution, a factor largely overlooked in prior work. We illustrate this with a simple yet effective adversarial attack, underscoring the need for a more systematic study.

stat.ML↗

arXiv2025

Sparse, Efficient and Explainable Data Attribution with DualXDA

Data Attribution (DA) is an emerging approach in the field of eXplainable Artificial Intelligence (XAI), aiming to identify influential training datapoints which determine model outputs. It seeks to provide transparency about the model and individual predictions, e.g. for model debugging, identifying data-related causes of suboptimal performance. However, existing DA approaches suffer from prohibitively high computational costs and memory demands when applied to even medium-scale datasets and models, forcing practitioners to resort to approximations that may fail to capture the true inference process of the underlying model. Additionally, current attribution methods exhibit low sparsity, resulting in non-negligible attribution scores across a high number of training examples, hindering the discovery of decisive patterns in the data. In this work, we introduce DualXDA, a framework for sparse, efficient and explainable DA, comprised of two interlinked approaches, Dual Data Attribution (DualDA) and eXplainable Data Attribution (XDA): With DualDA, we propose a novel approach for efficient and effective DA, leveraging Support Vector Machine theory to provide fast and naturally sparse data attributions for AI predictions. In extensive quantitative analyses, we demonstrate that DualDA achieves high attribution quality, excels at solving a series of evaluated downstream tasks, while at the same time improving explanation time by a factor of up to 4,100,000x compared to the original Influence Functions method, and up to 11,000x compared to the method's most efficient approximation from literature to date. We further introduce XDA, a method for enhancing Data Attribution with capabilities from feature attribution methods to explain why training samples are relevant for the prediction of a test sample in terms of impactful features, which we showcase and verify qualitatively in detail.

cs.LG↗

arXiv2025

Optimizing Federated Learning by Entropy-Based Client Selection

Although deep learning has revolutionized domains such as natural language processing and computer vision, its dependence on centralized datasets raises serious privacy concerns. Federated learning addresses this issue by enabling multiple clients to collaboratively train a global deep learning model without compromising their data privacy. However, the performance of such a model degrades under label skew, where the label distribution differs between clients. To overcome this issue, a novel method called FedEntOpt is proposed. In each round, it selects clients to maximize the entropy of the aggregated label distribution, ensuring that the global model is exposed to data from all available classes. Extensive experiments on multiple benchmark datasets show that the proposed method outperforms several state-of-the-art algorithms by up to 6% in classification accuracy under standard settings regardless of the model size, while achieving gains of over 30% in scenarios with low participation rates and client dropout. In addition, FedEntOpt offers the flexibility to be combined with existing algorithms, enhancing their classification accuracy by more than 40%. Importantly, its performance remains unaffected even when differential privacy is applied.

cs.LG↗

arXiv2025

Multimodal Deep Learning for Prediction of Progression-Free Survival in Patients with Neuroendocrine Tumors Undergoing 177Lu-based Peptide Receptor Radionuclide Therapy

Peptide receptor radionuclide therapy (PRRT) is an established treatment for metastatic neuroendocrine tumors (NETs), yet long-term disease control occurs only in a subset of patients. Predicting progression-free survival (PFS) could support individualized treatment planning. This study evaluates laboratory, imaging, and multimodal deep learning models for PFS prediction in PRRT-treated patients. In this retrospective, single-center study 116 patients with metastatic NETs undergoing 177Lu-DOTATOC were included. Clinical characteristics, laboratory values, and pretherapeutic somatostatin receptor positron emission tomography/computed tomographies (SR-PET/CT) were collected. Seven models were trained to classify low- vs. high-PFS groups, including unimodal (laboratory, SR-PET, or CT) and multimodal fusion approaches. Explainability was evaluated by feature importance analysis and gradient maps. Forty-two patients (36%) had short PFS (< 1 year), 74 patients long PFS (>1 year). Groups were similar in most characteristics, except for higher baseline chromogranin A (p = 0.003), elevated gamma-GT (p = 0.002), and fewer PRRT cycles (p < 0.001) in short-PFS patients. The Random Forest model trained only on laboratory biomarkers reached an AUROC of 0.59 +- 0.02. Unimodal three-dimensional convolutional neural networks using SR-PET or CT performed worse (AUROC 0.42 +- 0.03 and 0.54 +- 0.01, respectively). A multimodal fusion model laboratory values, SR-PET, and CT -augmented with a pretrained CT branch - achieved the best results (AUROC 0.72 +- 0.01, AUPRC 0.80 +- 0.01). Multimodal deep learning combining SR-PET, CT, and laboratory biomarkers outperformed unimodal approaches for PFS prediction after PRRT. Upon external validation, such models may support risk-adapted follow-up strategies.

cs.LG↗

arXiv2025

Fractional Diffusion Bridge Models

We present Fractional Diffusion Bridge Models (FDBM), a novel generative diffusion bridge framework driven by an approximation of the rich and non-Markovian fractional Brownian motion (fBM). Real stochastic processes exhibit a degree of memory effects (correlations in time), long-range dependencies, roughness and anomalous diffusion phenomena that are not captured in standard diffusion or bridge modeling due to the use of Brownian motion (BM). As a remedy, leveraging a recent Markovian approximation of fBM (MA-fBM), we construct FDBM that enable tractable inference while preserving the non-Markovian nature of fBM. We prove the existence of a coupling-preserving generative diffusion bridge and leverage it for future state prediction from paired training data. We then extend our formulation to the Schrödinger bridge problem and derive a principled loss function to learn the unpaired data translation. We evaluate FDBM on both tasks: predicting future protein conformations from aligned data, and unpaired image translation. In both settings, FDBM achieves superior performance compared to the Brownian baselines, yielding lower root mean squared deviation (RMSD) of C$_α$ atomic positions in protein structure prediction and lower Fréchet Inception Distance (FID) in unpaired image translation.

cs.LG↗

arXiv2025

The Atlas of In-Context Learning: How Attention Heads Shape In-Context Retrieval Augmentation

Large language models are able to exploit in-context learning to access external knowledge beyond their training data through retrieval-augmentation. While promising, its inner workings remain unclear. In this work, we shed light on the mechanism of in-context retrieval augmentation for question answering by viewing a prompt as a composition of informational components. We propose an attribution-based method to identify specialized attention heads, revealing in-context heads that comprehend instructions and retrieve relevant contextual information, and parametric heads that store entities' relational knowledge. To better understand their roles, we extract function vectors and modify their attention weights to show how they can influence the answer generation process. Finally, we leverage the gained insights to trace the sources of knowledge used during inference, paving the way towards more safe and transparent language models.

cs.CL↗

arXiv2025

Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are crafted to be subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using them as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

cs.LG↗