arXiv subjects

Sebastian Lapuschkin

Publications and source records attributed to Sebastian Lapuschkin.

At least 37 records · Page 2Linked to original sources

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

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

Manipulating Feature Visualizations with Gradient Slingshots

Feature Visualization (FV) is a widely used technique for interpreting concepts learned by Deep Neural Networks (DNNs), which synthesizes input patterns that maximally activate a given feature. Despite its popularity, the trustworthiness of FV explanations has received limited attention. We introduce Gradient Slingshots, a novel method that enables FV manipulation without modifying model architecture or significantly degrading performance. By shaping new trajectories in off-distribution regions of a feature's activation landscape, we coerce the optimization process to converge to a predefined visualization. We evaluate our approach on several DNN architectures, demonstrating its ability to replace faithful FVs with arbitrary targets. These results expose a critical vulnerability: auditors relying solely on FV may accept entirely fabricated explanations. To mitigate this risk, we propose a straightforward defense and quantitatively demonstrate its effectiveness.

cs.LG↗

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

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↗

arXiv2025

See What I Mean? CUE: A Cognitive Model of Understanding Explanations

As machine learning systems increasingly inform critical decisions, the need for human-understandable explanations grows. Current evaluations of Explainable AI (XAI) often prioritize technical fidelity over cognitive accessibility which critically affects users, in particular those with visual impairments. We propose CUE, a model for Cognitive Understanding of Explanations, linking explanation properties to cognitive sub-processes: legibility (perception), readability (comprehension), and interpretability (interpretation). In a study (N=455) testing heatmaps with varying colormaps (BWR, Cividis, Coolwarm), we found comparable task performance but lower confidence/effort for visually impaired users. Unlike expected, these gaps were not mitigated and sometimes worsened by accessibility-focused color maps like Cividis. These results challenge assumptions about perceptual optimization and support the need for adaptive XAI interfaces. They also validate CUE by demonstrating that altering explanation legibility affects understandability. We contribute: (1) a formalized cognitive model for explanation understanding, (2) an integrated definition of human-centered explanation properties, and (3) empirical evidence motivating accessible, user-tailored XAI.

cs.HC↗

arXiv2025

Ensuring Medical AI Safety: Interpretability-Driven Detection and Mitigation of Spurious Model Behavior and Associated Data

Deep neural networks are increasingly employed in high-stakes medical applications, despite their tendency for shortcut learning in the presence of spurious correlations, which can have potentially fatal consequences in practice. Whereas a multitude of works address either the detection or mitigation of such shortcut behavior in isolation, the Reveal2Revise approach provides a comprehensive bias mitigation framework combining these steps. However, effectively addressing these biases often requires substantial labeling efforts from domain experts. In this work, we review the steps of the Reveal2Revise framework and enhance it with semi-automated interpretability-based bias annotation capabilities. This includes methods for the sample- and feature-level bias annotation, providing valuable information for bias mitigation methods to unlearn the undesired shortcut behavior. We show the applicability of the framework using four medical datasets across two modalities, featuring controlled and real-world spurious correlations caused by data artifacts. We successfully identify and mitigate these biases in VGG16, ResNet50, and contemporary Vision Transformer models, ultimately increasing their robustness and applicability for real-world medical tasks. Our code is available at https://github.com/frederikpahde/medical-ai-safety.

cs.AI↗

arXiv2025

FADE: Why Bad Descriptions Happen to Good Features

Recent advances in mechanistic interpretability have highlighted the potential of automating interpretability pipelines in analyzing the latent representations within LLMs. While this may enhance our understanding of internal mechanisms, the field lacks standardized evaluation methods for assessing the validity of discovered features. We attempt to bridge this gap by introducing FADE: Feature Alignment to Description Evaluation, a scalable model-agnostic framework for automatically evaluating feature-to-description alignment. FADE evaluates alignment across four key metrics - Clarity, Responsiveness, Purity, and Faithfulness - and systematically quantifies the causes of the misalignment between features and their descriptions. We apply FADE to analyze existing open-source feature descriptions and assess key components of automated interpretability pipelines, aiming to enhance the quality of descriptions. Our findings highlight fundamental challenges in generating feature descriptions, particularly for SAEs compared to MLP neurons, providing insights into the limitations and future directions of automated interpretability. We release FADE as an open-source package at: https://github.com/brunibrun/FADE

cs.LG↗

arXiv2025

Efficient and Flexible Neural Network Training through Layer-wise Feedback Propagation

Gradient-based optimization has been a cornerstone of machine learning that enabled the vast advances of Artificial Intelligence (AI) development over the past decades. However, this type of optimization requires differentiation, and with recent evidence of the benefits of non-differentiable (e.g. neuromorphic) architectures over classical models w.r.t. efficiency, such constraints can become limiting in the future. We present Layer-wise Feedback Propagation (LFP), a novel training principle for neural network-like predictors that utilizes methods from the domain of explainability to decompose a reward to individual neurons based on their respective contributions. Leveraging these neuron-wise rewards, our method then implements a greedy approach reinforcing helpful parts of the network and weakening harmful ones. While having comparable computational complexity to gradient descent, LFP does not require gradient computation and generates sparse and thereby memory- and energy-efficient parameter updates and models. We establish the convergence of LFP theoretically and empirically, demonstrating its effectiveness on various models and datasets. Via two applications - neural network pruning and the approximation-free training of Spiking Neural Networks (SNNs) - we demonstrate that LFP combines increased efficiency in terms of computation and representation with flexibility w.r.t. choice of model architecture and objective function. Our code is available at https://github.com/leanderweber/layerwise-feedback-propagation.

cs.LG↗

arXiv2025

Leveraging Influence Functions for Resampling Data in Physics-Informed Neural Networks

Physics-informed neural networks (PINNs) offer a powerful approach to solving partial differential equations (PDEs), which are ubiquitous in the quantitative sciences. Applied to both forward and inverse problems across various scientific domains, PINNs have recently emerged as a valuable tool in the field of scientific machine learning. A key aspect of their training is that the data -- spatio-temporal points sampled from the PDE's input domain -- are readily available. Influence functions, a tool from the field of explainable AI (XAI), approximate the effect of individual training points on the model, enhancing interpretability. In the present work, we explore the application of influence function-based sampling approaches for the training data. Our results indicate that such targeted resampling based on data attribution methods has the potential to enhance prediction accuracy in physics-informed neural networks, demonstrating a practical application of an XAI method in PINN training.

cs.LG↗

arXiv2025

Deep Learning-based Multi Project InP Wafer Simulation for Unsupervised Surface Defect Detection

Quality management in semiconductor manufacturing often relies on template matching with known golden standards. For Indium-Phosphide (InP) multi-project wafer manufacturing, low production scale and high design variability lead to such golden standards being typically unavailable. Defect detection, in turn, is manual and labor-intensive. This work addresses this challenge by proposing a methodology to generate a synthetic golden standard using Deep Neural Networks, trained to simulate photo-realistic InP wafer images from CAD data. We evaluate various training objectives and assess the quality of the simulated images on both synthetic data and InP wafer photographs. Our deep-learning-based method outperforms a baseline decision-tree-based approach, enabling the use of a 'simulated golden die' from CAD plans in any user-defined region of a wafer for more efficient defect detection. We apply our method to a template matching procedure, to demonstrate its practical utility in surface defect detection.

cs.CV↗

arXiv2025

Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video

Robust tooling and publicly available pre-trained models have helped drive recent advances in mechanistic interpretability for language models. However, similar progress in vision mechanistic interpretability has been hindered by the lack of accessible frameworks and pre-trained weights. We present Prisma (Access the codebase here: https://github.com/Prisma-Multimodal/ViT-Prisma), an open-source framework designed to accelerate vision mechanistic interpretability research, providing a unified toolkit for accessing 75+ vision and video transformers; support for sparse autoencoder (SAE), transcoder, and crosscoder training; a suite of 80+ pre-trained SAE weights; activation caching, circuit analysis tools, and visualization tools; and educational resources. Our analysis reveals surprising findings, including that effective vision SAEs can exhibit substantially lower sparsity patterns than language SAEs, and that in some instances, SAE reconstructions can decrease model loss. Prisma enables new research directions for understanding vision model internals while lowering barriers to entry in this emerging field.

cs.CV↗

arXiv2025

Relevance-driven Input Dropout: an Explanation-guided Regularization Technique

Overfitting is a well-known issue extending even to state-of-the-art (SOTA) Machine Learning (ML) models, resulting in reduced generalization, and a significant train-test performance gap. Mitigation measures include a combination of dropout, data augmentation, weight decay, and other regularization techniques. Among the various data augmentation strategies, occlusion is a prominent technique that typically focuses on randomly masking regions of the input during training. Most of the existing literature emphasizes randomness in selecting and modifying the input features instead of regions that strongly influence model decisions. We propose Relevance-driven Input Dropout (RelDrop), a novel data augmentation method which selectively occludes the most relevant regions of the input, nudging the model to use other important features in the prediction process, thus improving model generalization through informed regularization. We further conduct qualitative and quantitative analyses to study how Relevance-driven Input Dropout (RelDrop) affects model decision-making. Through a series of experiments on benchmark datasets, we demonstrate that our approach improves robustness towards occlusion, results in models utilizing more features within the region of interest, and boosts inference time generalization performance. Our code is available at https://github.com/Shreyas-Gururaj/LRP_Relevance_Dropout.

cs.LG↗