arXiv subjects

Sebastian Lapuschkin

Publications and source records attributed to Sebastian Lapuschkin.

At least 19 recordsLinked to original sources

arXiv2026

Concept-based explanation of gene expression prediction from H&E images

Recent advances in pathology foundation models have enabled accurate prediction of spatial transcriptomics (ST) from routine H&E images. However, existing explainability methods for vision transformer (ViT)-based models are largely limited to local heatmaps and do not reveal how morphological concepts contribute to ST predictions. Here, we introduce an explainable framework that combines relevance propagation and concept discovery to link transcriptional programs to tissue morphology. We developed a ViT-based framework for virtual ST from H&E images that combines ViT-aware layer-wise relevance propagation with relaxed archetypal TopK sparse autoencoder-based concept discovery. This approach provides both local explanations and global insights into the morphological patterns associated with transcriptional programs. We applied the framework to colorectal cancer ST data from the HEST-1k cohort and evaluated its generalizability in TCGA COAD. Our architecture accurately predicts clinically relevant ST signatures and accompanying molecular phenotypes. Measured and predicted gene expression profiles reveal substantial spatial heterogeneity of the colorectal cancer subtypes iCMS2 and iCMS3 across a large number of samples. Spatially resolved and aggregated iCMS classification achieve weighted F1 scores of 0.872 and 0.819 (0.770 in TCGA COAD), respectively, and both stratify patient outcome. Beyond prediction, our framework establishes a relevance-based concept atlas linking molecular phenotypes to histopathological representations. Comparison of activation- with relevance-derived concepts demonstrates that relevances provide a more direct link between tissue morphology and downstream predictions. We establish a general strategy for concept-based explanation of spatial prediction, and our framework is readily applicable to a broad range of ViT-based pathology models.

cs.CV↗

arXiv2026

Predicting Future Behaviors in Reasoning Models Enables Better Steering

Deployed large reasoning models (LRMs) often behave unexpectedly. Test-time steering controls LRM outputs by intervening on their hidden representations, but it can degrade output quality. We argue that prior steering work implicitly relies on internal features that detect behavior in already generated text. We show that these detection features are poor predictors of future behavioral outcomes, and thus not the natural intervention target. Instead, we train activation probes to predict future behavior likelihoods from intermediate reasoning steps. These probes predict the most likely behavior with 64%-91% accuracy, revealing a separate type of internal prediction features. Building on these prediction features, we introduce a text-level steering method, Future Probe Controlled Generation. FPCG samples multiple candidate sentences and chooses the best one according to a probe predicting the future behavior likelihood. This enables steering with almost no output quality degradation. FPCG also enables steering in several evaluations where activation steering fails. These results show that distinguishing detection and prediction features enables a more nuanced approach to controlling LRM behaviors.

cs.LG↗

arXiv2026

Fast & Faithful Function Vectors

Function vectors (FVs) are task representations elicited during in-context learning that can be used to steer Large Language Models (LLMs). However, design choices in their formulation remain underexplored. In this work, we study the impact of varying FV definitions for instructions along two degrees of freedom: attention head selection and steering. For head selection, using gradient-based attributions with Layer-wise Relevance Propagation (LRP) substantially improves efficiency as well as accuracy. For FV steering, applying it in a distributed manner yields a higher accuracy compared to simple aggregation. Our code is publicly available.

cs.CL↗

arXiv2026

Playing the network backward: A Game Theoretic Attribution Framework

Attribution methods explain which input features drive a model's prediction, making them central to model debugging and mechanistic interpretability. Yet backward attribution methods, including gradients, LRP, and transformer-specific rules, lack a shared framework in which to compare the underlying backward calculations. We introduce such a framework by recasting backward attribution as a two-player game on an extended network graph, building on Gaubert and Vlassopoulos' ReLU Net Game. Gradients and the full alpha-beta-LRP family arise as integrals over game trajectories under specific equilibria, so attribution maps become projections of trajectory distributions rather than the primary object. Desired explanation properties, such as localisation focus, robustness to input noise, or stable attention routing, can be specified as game-theoretic concepts, including policy regularization, risk aversion, and extended action sets, and translate directly into novel adaptations of the well-known backward rules. On ViT-B/16, one such selected adaptation of alpha-beta-LRP outperforms prior transformer-specific backward methods across all considered localisation metrics.

cs.LG↗

arXiv2026

Contrastive Semantic Projection: Faithful Neuron Labeling with Contrastive Examples

Neuron labeling assigns textual descriptions to internal units of deep networks. Existing approaches typically rely on highly activating examples, often yielding broad or misleading labels by focusing on dominant but incidental visual factors. Prior work such as FALCON introduced contrastive examples -- inputs that are semantically similar to activating examples but elicit low activations -- to sharpen explanations, but it primarily addresses subspace-level interpretability rather than scalable neuron-level labeling. We revisit contrastive explanations for neuron-level labeling in two stages: (1) candidate label generation with vision language models (VLMs) and (2) label assignment with CLIP-like encoders. First, we show that providing contrastive image sets to VLMs yields candidate labels that are more specific and more faithful. Second, we introduce Contrastive Semantic Projection (CSP), an extension of SemanticLens that incorporates contrastive examples directly into its CLIP-based scoring and selection pipeline. Across extensive experiments and a case study on melanoma detection, contrastive labeling improves both faithfulness and semantic granularity over state-of-the-art baselines. Our results demonstrate that contrastive examples are a simple yet powerful and currently underutilized component of neuron labeling and analysis pipelines.

cs.CV↗

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

From Attribution to Action: A Human-Centered Application of Activation Steering

Explainable AI (XAI) methods reveal which features influence model predictions, yet provide limited means for practitioners to act on these explanations. Activation steering of components identified via XAI offers a path toward actionable explanations, although its practical utility remains understudied. We introduce an interactive workflow combining SAE-based attribution with activation steering for instance-level analysis of concept usage in vision models, implemented as a web-based tool. Based on this workflow, we conduct semi-structured expert interviews (N=8) with debugging tasks on CLIP to investigate how practitioners reason about, trust, and apply activation steering. We find that steering enables a shift from inspection to intervention-based hypothesis testing (8/8 participants), with most grounding trust in observed model responses rather than explanation plausibility alone (6/8). Participants adopted systematic debugging strategies dominated by component suppression (7/8) and highlighted risks including ripple effects and limited generalization of instance-level corrections. Overall, activation steering renders interpretability more actionable while raising important considerations for safe and effective use.

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

Concept-based explanations of Segmentation and Detection models in Natural Disaster Management

Deep learning models for flood and wildfire segmentation and object detection enable precise, real-time disaster localization when deployed on embedded drone platforms. However, in natural disaster management, the lack of transparency in their decision-making process hinders human trust required for emergency response. To address this, we present an explainability framework for understanding flood segmentation and car detection predictions on the widely used PIDNet and YOLO architectures. More specifically, we introduce a novel redistribution strategy that extends Layer-wise Relevance Propagation (LRP) explanations for sigmoid-gated element-wise fusion layers. This extension allows LRP relevances to flow through the fusion modules of PIDNet, covering the entire computation graph back to the input image. Furthermore, we apply Prototypical Concept-based Explanations (PCX) to provide both local and global explanations at the concept level, revealing which learned features drive the segmentation and detection of specific disaster semantic classes. Experiments on a publicly available flood dataset show that our framework provides reliable and interpretable explanations while maintaining near real-time inference capabilities, rendering it suitable for deployment on resource-constrained platforms, such as Unmanned Aerial Vehicles (UAVs).

cs.CV↗

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

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↗

arXiv2025

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↗

arXiv2025

LieSolver: PDE-Constrained Learning for IBVPs via Lie Symmetries

Initial-boundary value problems (IBVPs) provide the essential framework for modelling a wide range of phenomena in physics and engineering. We introduce a novel method for efficiently solving IBVPs using Lie symmetries to enforce the associated partial differential equation (PDE) exactly by construction. By leveraging symmetry transformations, our model embeds the underlying physical laws and learns the solution solely from initial and boundary data. Consequently, the boundary loss directly quantifies domain-wide error, enabling rigorous error estimation for well-posed IBVPs. We implement LieSolver and demonstrate its application to linear homogeneous PDEs, showing that it outperforms physics-informed neural networks (PINNs) in both speed and accuracy while yielding compact models. Overall, our approach significantly enhances the efficiency and reliability of predictions for PDE-constrained problems.

cs.LG↗

arXiv2025

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↗

arXiv2025

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↗

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↗

arXiv2025

The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network

Recent mechanistic work has uncovered learned algorithms within neural networks, from modular arithmetic to search and planning in game-playing agents. But does algorithmic structure guarantee algorithmic behavior? We investigate this in Leela Chess Zero, the strongest neural chess engine, where prior work identified learned look-ahead. By extending the logit lens to its move-selecting policy network, we discover that correct puzzle solutions-including immediate checkmates-often appear in intermediate layers but are systematically overridden in the final output, a phenomenon we term "forgotten puzzles". Replicating prior analyses on these positions, we find that look-ahead operates normally-future moves of the correct continuation are represented, causally important, and linearly decodable-ruling out a failure of the algorithm itself. Instead, late layers increasingly shift toward prioritizing safe play over aggression. To test whether this shift drives the override, we steer the model against these preferences and recover 61.7% of forgotten puzzles, providing causal evidence that safety priors override algorithmically computed solutions. These findings demonstrate that algorithmic structure does not guarantee algorithmic behavior: a model can internally solve a problem and still output the wrong answer.

cs.LG↗

arXiv2025

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↗