Selected work

Representative Papers

Towards Hierarchical Structure Understanding of Newspaper Images

Jul 16, 2026

This study addresses the challenges of document understanding in historical newspaper images, which exhibit complex nested structures and densely heterogeneous layouts. To tackle this problem, the authors propose two complementary approaches: a modular bottom-up pipeline integrating YOLO-based layout detection, LayoutReader for reading order prediction, and a custom article segmentation algorithm; and Tiramisu, a novel end-to-end hierarchical Transformer architecture that explicitly models the multi-level structure of newspapers through an iterative hierarchical mechanism. The work also contributes Finlam La Liberté, the first dataset tailored for hierarchical information retrieval in historical newspapers, along with open-sourced training code and a synthetic newspaper generator. Experimental results demonstrate that both methods effectively reconstruct intricate newspaper hierarchies, offering significant advantages for scalable document digitization.

0 citationsRead paper

Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets

Jul 05, 2026

This work addresses the challenge of identifying optimal adjustment sets in data-scarce settings, where local causal discovery is hindered by insufficient samples, incomplete neighborhoods, and unresolved Markov equivalence classes. To overcome these limitations, the authors propose b-LOAD, a novel method that uniquely integrates structured prior knowledge directly into the local causal discovery process. By leveraging constrained edge information and dynamically expanding the local graph boundary via Meek rules, b-LOAD constructs a knowledge-constrained local partially directed acyclic graph. This approach monotonically refines the admissible equivalence class, substantially broadening the scope of identifiable causal queries and recovering optimal adjustment sets that are otherwise unidentifiable from observational data alone. Experimental results demonstrate that b-LOAD significantly outperforms purely data-driven and conventional knowledge-enhanced baselines under data scarcity and structural complexity, with validation on real biological networks confirming its efficacy and robustness.

0 citationsRead paper

Scaling State-Space Models from Lines to Paragraphs: An Ablation of Mamba-based OCR

Jun 22, 2026

This study addresses the challenge of efficiently scaling optical character recognition (OCR) from single-line to paragraph-level text, where traditional Transformers suffer from quadratic complexity due to their attention mechanism, and the applicability of state space models like Mamba to long sequences and handwritten text remains unclear. The work presents the first systematic investigation of Mamba as an autoregressive decoder in OCR, employing controlled ablation studies to analyze the impact of key hyperparameters such as state dimension and expansion factor. Results show that on synthetic paragraph data, Mamba achieves a character error rate (CER) below 1% and inference speeds 1.4–4.5× faster than Transformer. However, on real-world handwritten data (IAM), its performance lags significantly (CER: 10.0% vs. 3.5%), indicating that the primary bottleneck stems from data scarcity rather than architectural limitations.

0 citationsRead paper

NoduLoCC2026: Lung Nodule Localization and Classification Contest from Chest X-Ray Images

Jun 19, 2026

This study addresses the insufficient accuracy in detecting and localizing pulmonary nodules in chest X-ray images by organizing an international challenge that systematically evaluates diverse deep learning approaches on a unified real-world dataset. The assessment encompasses object detection models, classification architectures, and post-processing strategies, revealing inherent difficulties in nodule localization. Results demonstrate that the best-performing classification method achieves a balanced accuracy of 0.72 and an AUC-ROC of 0.79. However, for localization, only 53% of images yield correct predictions of nodule count, with a median distance error of 12.83 mm, underscoring the current limitations of existing methods in achieving precise spatial localization.

0 citationsRead paper

Beyond Dark Knowledge: Mixup-Based Distillation for Reliable Predictions

Jun 10, 2026

This work addresses a critical yet overlooked issue in knowledge distillation when combined with mixup: the teacher model is queried on out-of-distribution neighborhoods, leading to supervision signals corrupted by distributional confusion and impairing knowledge transfer. The authors propose a distillation mechanism that applies mixup exclusively to the student, thereby revealing for the first time the distributional mismatch between teacher and student. They demonstrate that calibration capability can be transferred independently of accuracy. By integrating temperature scaling with calibration-aware evaluation, the method’s universality across teachers of varying capacities is validated on CIFAR and ImageNet. Results show that the student not only achieves significantly higher accuracy but also exhibits an order-of-magnitude reduction in overconfidence, along with improved uncertainty estimation and representation geometry.

0 citationsRead paper

Recent publications

Latest Papers

Towards Hierarchical Structure Understanding of Newspaper Images

Jul 16, 2026

This study addresses the challenges of document understanding in historical newspaper images, which exhibit complex nested structures and densely heterogeneous layouts. To tackle this problem, the authors propose two complementary approaches: a modular bottom-up pipeline integrating YOLO-based layout detection, LayoutReader for reading order prediction, and a custom article segmentation algorithm; and Tiramisu, a novel end-to-end hierarchical Transformer architecture that explicitly models the multi-level structure of newspapers through an iterative hierarchical mechanism. The work also contributes Finlam La Liberté, the first dataset tailored for hierarchical information retrieval in historical newspapers, along with open-sourced training code and a synthetic newspaper generator. Experimental results demonstrate that both methods effectively reconstruct intricate newspaper hierarchies, offering significant advantages for scalable document digitization.

0 citationsRead paper

Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets

Jul 05, 2026

This work addresses the challenge of identifying optimal adjustment sets in data-scarce settings, where local causal discovery is hindered by insufficient samples, incomplete neighborhoods, and unresolved Markov equivalence classes. To overcome these limitations, the authors propose b-LOAD, a novel method that uniquely integrates structured prior knowledge directly into the local causal discovery process. By leveraging constrained edge information and dynamically expanding the local graph boundary via Meek rules, b-LOAD constructs a knowledge-constrained local partially directed acyclic graph. This approach monotonically refines the admissible equivalence class, substantially broadening the scope of identifiable causal queries and recovering optimal adjustment sets that are otherwise unidentifiable from observational data alone. Experimental results demonstrate that b-LOAD significantly outperforms purely data-driven and conventional knowledge-enhanced baselines under data scarcity and structural complexity, with validation on real biological networks confirming its efficacy and robustness.

0 citationsRead paper

Scaling State-Space Models from Lines to Paragraphs: An Ablation of Mamba-based OCR

Jun 22, 2026

This study addresses the challenge of efficiently scaling optical character recognition (OCR) from single-line to paragraph-level text, where traditional Transformers suffer from quadratic complexity due to their attention mechanism, and the applicability of state space models like Mamba to long sequences and handwritten text remains unclear. The work presents the first systematic investigation of Mamba as an autoregressive decoder in OCR, employing controlled ablation studies to analyze the impact of key hyperparameters such as state dimension and expansion factor. Results show that on synthetic paragraph data, Mamba achieves a character error rate (CER) below 1% and inference speeds 1.4–4.5× faster than Transformer. However, on real-world handwritten data (IAM), its performance lags significantly (CER: 10.0% vs. 3.5%), indicating that the primary bottleneck stems from data scarcity rather than architectural limitations.

0 citationsRead paper

NoduLoCC2026: Lung Nodule Localization and Classification Contest from Chest X-Ray Images

Jun 19, 2026

This study addresses the insufficient accuracy in detecting and localizing pulmonary nodules in chest X-ray images by organizing an international challenge that systematically evaluates diverse deep learning approaches on a unified real-world dataset. The assessment encompasses object detection models, classification architectures, and post-processing strategies, revealing inherent difficulties in nodule localization. Results demonstrate that the best-performing classification method achieves a balanced accuracy of 0.72 and an AUC-ROC of 0.79. However, for localization, only 53% of images yield correct predictions of nodule count, with a median distance error of 12.83 mm, underscoring the current limitations of existing methods in achieving precise spatial localization.

0 citationsRead paper

Beyond Dark Knowledge: Mixup-Based Distillation for Reliable Predictions

Jun 10, 2026

This work addresses a critical yet overlooked issue in knowledge distillation when combined with mixup: the teacher model is queried on out-of-distribution neighborhoods, leading to supervision signals corrupted by distributional confusion and impairing knowledge transfer. The authors propose a distillation mechanism that applies mixup exclusively to the student, thereby revealing for the first time the distributional mismatch between teacher and student. They demonstrate that calibration capability can be transferred independently of accuracy. By integrating temperature scaling with calibration-aware evaluation, the method’s universality across teachers of varying capacities is validated on CIFAR and ImageNet. Results show that the student not only achieves significantly higher accuracy but also exhibits an order-of-magnitude reduction in overconfidence, along with improved uncertainty estimation and representation geometry.

0 citationsRead paper