arXiv subjects

Dimosthenis Karatzas

Publications and source records attributed to Dimosthenis Karatzas.

At least 19 recordsLinked to original sources

arXiv2026

Rethinking Expert Training for Model Merging with Prompt Learning

Model merging aims to combine multiple domain-specialized experts trained from a shared foundation model into a single multi-task model. Existing approaches largely focus on improving the merging procedure itself and typically assume experts obtained through full-parameter fine-tuning. In this work, we revisit expert training for model merging. We first show that prompt-based adaptation provides a strong baseline: independently learned prompts can be exploited across tasks while keeping the backbone fixed, avoiding the interference introduced by weight merging. Building on this observation, we introduce Dual-Tuned Experts (DTEs), a two-stage training strategy that first learns prompts and then fine-tunes the vision encoder. This reduces the magnitude of task-specific parameter updates and produces experts with higher merge compatibility. Experiments across multiple CLIP architectures, full fine-tuning, and LoRA experts show that DTEs consistently improve merged performance of standard merging approaches and remain effective even when combining heterogeneous sets of experts.

cs.CV↗

arXiv2026

SMI: Efficient Self-Supervised Learning via Mutual-Information-Inspired Dependency Optimization

Self-supervised learning (SSL) has achieved remarkable representation learning performance, but many existing methods rely on large batch sizes, memory banks, momentum encoders, or global synchronization mechanisms that substantially increase computational cost and training complexity. In this work, we propose Semantic Mutual Information (SMI), a lightweight self-supervised objective derived from a mutual-information-inspired dependency formulation under Gaussian assumptions. Unlike conventional correlation matching objectives that operate on high-dimensional feature correlation matrices, SMI performs optimization on a sample-level dependency matrix through a nonlinear transformation of pairwise correlations. This formulation induces distinct optimization dynamics that emphasize strongly dependent semantic pairs while maintaining representation diversity. Experimental results on ImageNet using a ResNet-50 backbone demonstrate that SMI achieves competitive linear evaluation performance relative to state-of-the-art SSL approaches while substantially reducing computational complexity. Across multiple low-resource benchmarks, SMI consistently improves transfer performance over Barlow Twins, particularly on fine-grained datasets. Furthermore, analyses of optimization dynamics and representation geometry suggest improved alignment--redundancy balance, greater feature diversity, and more spatially localized semantic representations. These results indicate that nonlinear dependency optimization provides an effective and computationally efficient alternative to conventional correlation-based self-supervised learning objectives.

cs.CV↗

arXiv2026

Learning Quantifiable Visual Explanations Without Ground-Truth

Explainable AI (XAI) techniques are increasingly important for the validation and responsible use of modern deep learning models, but are difficult to evaluate due to the lack of good ground-truth to compare against. We propose a framework that serves as a quantifiable metric for the quality of XAI methods, based on continuous input perturbation. Our metric formally considers the sufficiency and necessity of the attributed information to the model's decision-making, and we illustrate a range of cases where it aligns better with human intuitions of explanation quality than do existing metrics. To exploit the properties of this metric, we also propose a novel XAI method, considering the case where we fine-tune a model using a differentiable approximation of the metric as a supervision signal. The result is an adapter module that can be trained on top of any black-box model to output causal explanations of the model's decision process, without degrading model performance. We show that the explanations generated by this method outperform those of competing XAI techniques according to a number of quantifiable metrics.

cs.AI↗

arXiv2026

TRIMMER: A New Paradigm for Video Summarization through Self-Supervised Reinforcement Learning

The rapid growth of video content across domains such as surveillance, education, and social media has made efficient content understanding increasingly critical. Video summarization addresses this challenge by generating concise yet semantically meaningful representations, but existing approaches often rely on expensive manual annotations, struggle to generalize across domains, and incur significant computational costs due to complex architectures. Moreover, unsupervised and weakly supervised methods typically underperform compared to supervised counterparts in capturing long-range temporal dependencies and semantic structure. In this work, we propose TRIMMER (Temporal Relative Information Maximization for Multi-objective Efficient Reinforcement), a novel self-supervised reinforcement learning framework for video summarization. TRIMMER operates in two stages: it first learns robust representations via self-supervised learning and then performs spatio-temporal decision making through reinforcement learning guided by information-theoretic reward functions. Unlike prior approaches that rely on similarity-based objectives, our method introduces entropy-based metrics to capture higher-order temporal dynamics and semantic diversity, while computing rewards directly over selected frame indices to improve computational efficiency. Extensive experiments on standard benchmarks demonstrate that TRIMMER achieves state-of-the-art performance among unsupervised and self-supervised methods, while remaining competitive with leading supervised approaches, highlighting its effectiveness for scalable and generalizable video summarization.

cs.CV↗

arXiv2026

Reading in the Dark: Low-light Scene Text Recognition

Accurate text recognition in low-light environments is essential for intelligent systems in applications ranging from autonomous vehicles to smart surveillance. However, challenges such as poor illumination and noise interference remain underexplored. To address this gap, we introduce LSTR, a large-scale Low-light Scene Text Recognition dataset comprising 11,273 low-light images generated from well-lit datasets (ICDAR2015, IIIT5K, and WordArt), along with ESTR, which includes 60 real nighttime street-scene images in English and Spanish for exclusive evaluation. We explore two solution strategies: (1) employing Optical Character Recognition (OCR) models with fine-tuning and LoRA-based fine-tuning and (2) a joint training strategy that integrates a low-light image enhancement (LLIE) module with an OCR model. In particular, we propose a novel re-render LLIE (RLLIE) module, which demonstrates improved performance on real-world data. Through extensive experimentation, we analyze various training strategies and address a key research question: \emph{How bright is bright enough for effective scene text recognition?} Our results indicate that standalone LLIE or OCR models perform inadequately under low-light conditions, highlighting the advantages of specialized, jointly trained text-centric approaches. Additionally, we provide a comprehensive benchmark to support future research in robust low-light scene text recognition. https://huggingface.co/datasets/lumimusta/Low-light_Scene_Text_Dataset.

cs.CV↗

arXiv2026

Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections

Multimodal agents offer a promising path to automating complex document-intensive workflows. Yet, a critical question remains: do these agents demonstrate genuine strategic reasoning, or merely stochastic trial-and-error search? To address this, we introduce MADQA, a benchmark of 2,250 human-authored questions grounded in 800 heterogeneous PDF documents. Guided by Classical Test Theory, we design it to maximize discriminative power across varying levels of agentic abilities. To evaluate agentic behaviour, we introduce a novel evaluation protocol measuring the accuracy-effort trade-off. Using this framework, we show that while the best agents can match human searchers in raw accuracy, they succeed on largely different questions and rely on brute-force search to compensate for weak strategic planning. They fail to close the nearly 20% gap to oracle performance, persisting in unproductive loops. We release the dataset and evaluation harness to help facilitate the transition from brute-force retrieval to calibrated, efficient reasoning.

cs.CL↗

arXiv2026

AVIR: Adaptive Visual In-Document Retrieval for Efficient Multi-Page Document Question Answering

Multi-page Document Visual Question Answering (MP-DocVQA) remains challenging because long documents not only strain computational resources but also reduce the effectiveness of the attention mechanism in large vision-language models (LVLMs). We tackle these issues with an Adaptive Visual In-document Retrieval (AVIR) framework. A lightweight retrieval model first scores each page for question relevance. Pages are then clustered according to the score distribution to adaptively select relevant content. The clustered pages are screened again by Top-K to keep the context compact. However, for short documents, clustering reliability decreases, so we use a relevance probability threshold to select pages. The selected pages alone are fed to a frozen LVLM for answer generation, eliminating the need for model fine-tuning. The proposed AVIR framework reduces the average page count required for question answering by 70%, while achieving an ANLS of 84.58% on the MP-DocVQA dataset-surpassing previous methods with significantly lower computational cost. The effectiveness of the proposed AVIR is also verified on the SlideVQA and DUDE benchmarks. The code is available at https://github.com/Li-yachuan/AVIR.

cs.CV↗

arXiv2025

Counterfeit Answers: Adversarial Forgery against OCR-Free Document Visual Question Answering

Document Visual Question Answering (DocVQA) enables end-to-end reasoning grounded on information present in a document input. While recent models have shown impressive capabilities, they remain vulnerable to adversarial attacks. In this work, we introduce a novel attack scenario that aims to forge document content in a visually imperceptible yet semantically targeted manner, allowing an adversary to induce specific or generally incorrect answers from a DocVQA model. We develop specialized attack algorithms that can produce adversarially forged documents tailored to different attackers' goals, ranging from targeted misinformation to systematic model failure scenarios. We demonstrate the effectiveness of our approach against two end-to-end state-of-the-art models: Pix2Struct, a vision-language transformer that jointly processes image and text through sequence-to-sequence modeling, and Donut, a transformer-based model that directly extracts text and answers questions from document images. Our findings highlight critical vulnerabilities in current DocVQA systems and call for the development of more robust defenses. We release our open source code at https://github.com/pralab/adv-docVQA.

cs.CV↗

arXiv2025

Spatially Grounded Explanations in Vision Language Models for Document Visual Question Answering

We introduce EaGERS, a fully training-free and model-agnostic pipeline that (1) generates natural language rationales via a vision language model, (2) grounds these rationales to spatial sub-regions by computing multimodal embedding similarities over a configurable grid with majority voting, and (3) restricts the generation of responses only from the relevant regions selected in the masked image. Experiments on the DocVQA dataset demonstrate that our best configuration not only outperforms the base model on exact match accuracy and Average Normalized Levenshtein Similarity metrics but also enhances transparency and reproducibility in DocVQA without additional model fine-tuning.

cs.CV↗

arXiv2025

CoSMo: A Multimodal Transformer for Page Stream Segmentation in Comic Books

This paper introduces CoSMo, a novel multimodal Transformer for Page Stream Segmentation (PSS) in comic books, a critical task for automated content understanding, as it is a necessary first stage for many downstream tasks like character analysis, story indexing, or metadata enrichment. We formalize PSS for this unique medium and curate a new 20,800-page annotated dataset. CoSMo, developed in vision-only and multimodal variants, consistently outperforms traditional baselines and significantly larger general-purpose vision-language models across F1-Macro, Panoptic Quality, and stream-level metrics. Our findings highlight the dominance of visual features for comic PSS macro-structure, yet demonstrate multimodal benefits in resolving challenging ambiguities. CoSMo establishes a new state-of-the-art, paving the way for scalable comic book analysis.

cs.CV↗

arXiv2025

TRIM: A Self-Supervised Video Summarization Framework Maximizing Temporal Relative Information and Representativeness

The increasing ubiquity of video content and the corresponding demand for efficient access to meaningful information have elevated video summarization and video highlights as a vital research area. However, many state-of-the-art methods depend heavily either on supervised annotations or on attention-based models, which are computationally expensive and brittle in the face of distribution shifts that hinder cross-domain applicability across datasets. We introduce a pioneering self-supervised video summarization model that captures both spatial and temporal dependencies without the overhead of attention, RNNs, or transformers. Our framework integrates a novel set of Markov process-driven loss metrics and a two-stage self supervised learning paradigm that ensures both performance and efficiency. Our approach achieves state-of-the-art performance on the SUMME and TVSUM datasets, outperforming all existing unsupervised methods. It also rivals the best supervised models, demonstrating the potential for efficient, annotation-free architectures. This paves the way for more generalizable video summarization techniques and challenges the prevailing reliance on complex architectures.

cs.CV↗

arXiv2025

LLM-Driven Medical Document Analysis: Enhancing Trustworthy Pathology and Differential Diagnosis

Medical document analysis plays a crucial role in extracting essential clinical insights from unstructured healthcare records, supporting critical tasks such as differential diagnosis. Determining the most probable condition among overlapping symptoms requires precise evaluation and deep medical expertise. While recent advancements in large language models (LLMs) have significantly enhanced performance in medical document analysis, privacy concerns related to sensitive patient data limit the use of online LLMs services in clinical settings. To address these challenges, we propose a trustworthy medical document analysis platform that fine-tunes a LLaMA-v3 using low-rank adaptation, specifically optimized for differential diagnosis tasks. Our approach utilizes DDXPlus, the largest benchmark dataset for differential diagnosis, and demonstrates superior performance in pathology prediction and variable-length differential diagnosis compared to existing methods. The developed web-based platform allows users to submit their own unstructured medical documents and receive accurate, explainable diagnostic results. By incorporating advanced explainability techniques, the system ensures transparent and reliable predictions, fostering user trust and confidence. Extensive evaluations confirm that the proposed method surpasses current state-of-the-art models in predictive accuracy while offering practical utility in clinical settings. This work addresses the urgent need for reliable, explainable, and privacy-preserving artificial intelligence solutions, representing a significant advancement in intelligent medical document analysis for real-world healthcare applications. The code can be found at \href{https://github.com/leitro/Differential-Diagnosis-LoRA}{https://github.com/leitro/Differential-Diagnosis-LoRA}.

cs.AI↗

arXiv2025

DocVXQA: Context-Aware Visual Explanations for Document Question Answering

We propose DocVXQA, a novel framework for visually self-explainable document question answering. The framework is designed not only to produce accurate answers to questions but also to learn visual heatmaps that highlight contextually critical regions, thereby offering interpretable justifications for the model's decisions. To integrate explanations into the learning process, we quantitatively formulate explainability principles as explicit learning objectives. Unlike conventional methods that emphasize only the regions pertinent to the answer, our framework delivers explanations that are \textit{contextually sufficient} while remaining \textit{representation-efficient}. This fosters user trust while achieving a balance between predictive performance and interpretability in DocVQA applications. Extensive experiments, including human evaluation, provide strong evidence supporting the effectiveness of our method. The code is available at https://github.com/dali92002/DocVXQA.

cs.CV↗

arXiv2025

xLSTM-ECG: Multi-label ECG Classification via Feature Fusion with xLSTM

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, highlighting the critical need for efficient and accurate diagnostic tools. Electrocardiograms (ECGs) are indispensable in diagnosing various heart conditions; however, their manual interpretation is time-consuming and error-prone. In this paper, we propose xLSTM-ECG, a novel approach that leverages an extended Long Short-Term Memory (xLSTM) network for multi-label classification of ECG signals, using the PTB-XL dataset. To the best of our knowledge, this work represents the first design and application of xLSTM modules specifically adapted for multi-label ECG classification. Our method employs a Short-Time Fourier Transform (STFT) to convert time-series ECG waveforms into the frequency domain, thereby enhancing feature extraction. The xLSTM architecture is specifically tailored to address the complexities of 12-lead ECG recordings by capturing both local and global signal features. Comprehensive experiments on the PTB-XL dataset reveal that our model achieves strong multi-label classification performance, while additional tests on the Georgia 12-Lead dataset underscore its robustness and efficiency. This approach significantly improves ECG classification accuracy, thereby advancing clinical diagnostics and patient care. The code will be publicly available upon acceptance.

eess.SP↗

arXiv2025

NoTeS-Bank: Benchmarking Neural Transcription and Search for Scientific Notes Understanding

Understanding and reasoning over academic handwritten notes remains a challenge in document AI, particularly for mathematical equations, diagrams, and scientific notations. Existing visual question answering (VQA) benchmarks focus on printed or structured handwritten text, limiting generalization to real-world note-taking. To address this, we introduce NoTeS-Bank, an evaluation benchmark for Neural Transcription and Search in note-based question answering. NoTeS-Bank comprises complex notes across multiple domains, requiring models to process unstructured and multimodal content. The benchmark defines two tasks: (1) Evidence-Based VQA, where models retrieve localized answers with bounding-box evidence, and (2) Open-Domain VQA, where models classify the domain before retrieving relevant documents and answers. Unlike classical Document VQA datasets relying on optical character recognition (OCR) and structured data, NoTeS-BANK demands vision-language fusion, retrieval, and multimodal reasoning. We benchmark state-of-the-art Vision-Language Models (VLMs) and retrieval frameworks, exposing structured transcription and reasoning limitations. NoTeS-Bank provides a rigorous evaluation with NDCG@5, MRR, Recall@K, IoU, and ANLS, establishing a new standard for visual document understanding and reasoning.

cs.CV↗

arXiv2025

Preserving Privacy Without Compromising Accuracy: Machine Unlearning for Handwritten Text Recognition

Handwritten Text Recognition (HTR) is crucial for document digitization, but handwritten data can contain user-identifiable features, like unique writing styles, posing privacy risks. Regulations such as the ``right to be forgotten'' require models to remove these sensitive traces without full retraining. We introduce a practical encoder-only transformer baseline as a robust reference for future HTR research. Building on this, we propose a two-stage unlearning framework for multihead transformer HTR models. Our method combines neural pruning with machine unlearning applied to a writer classification head, ensuring sensitive information is removed while preserving the recognition head. We also present Writer-ID Confusion (WIC), a method that forces the forget set to follow a uniform distribution over writer identities, unlearning user-specific cues while maintaining text recognition performance. We compare WIC to Random Labeling, Fisher Forgetting, Amnesiac Unlearning, and DELETE within our prune-unlearn pipeline and consistently achieve better privacy and accuracy trade-offs. This is the first systematic study of machine unlearning for HTR. Using metrics such as Accuracy, Character Error Rate (CER), Word Error Rate (WER), and Membership Inference Attacks (MIA) on the IAM and CVL datasets, we demonstrate that our method achieves state-of-the-art or superior performance for effective unlearning. These experiments show that our approach effectively safeguards privacy without compromising accuracy, opening new directions for document analysis research. Our code is publicly available at https://github.com/leitro/WIC-WriterIDConfusion-MachineUnlearning.

cs.CV↗

arXiv2025

ComicsPAP: understanding comic strips by picking the correct panel

Large multimodal models (LMMs) have made impressive strides in image captioning, VQA, and video comprehension, yet they still struggle with the intricate temporal and spatial cues found in comics. To address this gap, we introduce ComicsPAP, a large-scale benchmark designed for comic strip understanding. Comprising over 100k samples and organized into 5 subtasks under a Pick-a-Panel framework, ComicsPAP demands models to identify the missing panel in a sequence. Our evaluations, conducted under both multi-image and single-image protocols, reveal that current state-of-the-art LMMs perform near chance on these tasks, underscoring significant limitations in capturing sequential and contextual dependencies. To close the gap, we adapted LMMs for comic strip understanding, obtaining better results on ComicsPAP than 10x bigger models, demonstrating that ComicsPAP offers a robust resource to drive future research in multimodal comic comprehension.

cs.CV↗

arXiv2025

DocMIA: Document-Level Membership Inference Attacks against DocVQA Models

Document Visual Question Answering (DocVQA) has introduced a new paradigm for end-to-end document understanding, and quickly became one of the standard benchmarks for multimodal LLMs. Automating document processing workflows, driven by DocVQA models, presents significant potential for many business sectors. However, documents tend to contain highly sensitive information, raising concerns about privacy risks associated with training such DocVQA models. One significant privacy vulnerability, exploited by the membership inference attack, is the possibility for an adversary to determine if a particular record was part of the model's training data. In this paper, we introduce two novel membership inference attacks tailored specifically to DocVQA models. These attacks are designed for two different adversarial scenarios: a white-box setting, where the attacker has full access to the model architecture and parameters, and a black-box setting, where only the model's outputs are available. Notably, our attacks assume the adversary lacks access to auxiliary datasets, which is more realistic in practice but also more challenging. Our unsupervised methods outperform existing state-of-the-art membership inference attacks across a variety of DocVQA models and datasets, demonstrating their effectiveness and highlighting the privacy risks in this domain.

cs.LG↗