Selected work

Representative Papers

Mind the Gap: Quantifying the Domain Gap in Cross-Sensor Diffusion Super-Resolution

Jun 26, 2026

This study addresses the domain gap between synthetic training data and real-world remote sensing imagery in cross-sensor super-resolution. It presents the first systematic evaluation of diffusion models on real Sentinel-2 and PlanetScope images, uncovering critical challenges in synthetic-to-real domain transfer. Leveraging a large-scale geometrically and temporally aligned dataset, the work integrates five state-of-the-art diffusion architectures with self-supervised learning and introduces LPIPS-Sat, a novel perceptual metric tailored for quantifying the domain discrepancy in satellite imagery. The findings reveal that models trained on synthetic data suffer significant performance degradation when applied to real scenes, while those trained directly on real data encounter optimization difficulties and struggle to generalize across the physical and radiometric diversity inherent in actual satellite observations, thereby highlighting fundamental limitations of current approaches.

0 citationsRead paper

V4FinBench: Benchmarking Tabular Foundation Models, LLMs, and Standard Methods on Corporate Bankruptcy Prediction

May 11, 2026

This study addresses the limitations of existing public datasets for corporate bankruptcy prediction—namely, their small scale, extreme class imbalance, and lack of a unified multi-horizon evaluation benchmark—by introducing V4FinBench, the first large-scale, publicly available, cross-regional (covering four countries in China and Europe) benchmark. V4FinBench incorporates 131 features, six prediction horizons, and a composite default definition. Using this benchmark, the authors systematically evaluate tabular foundation models (e.g., TabPFN), Llama-3-8B fine-tuned with QLoRA, and traditional gradient boosting methods under imbalance-aware training protocols. Results show that domain-adapted TabPFN matches or exceeds gradient boosting models in F1 and ROC-AUC for long-term predictions and demonstrates strong cross-regional generalization, whereas Llama-3-8B exhibits consistently weaker performance, particularly over longer horizons.

0 citationsRead paper

Unifying Deep Stochastic Processes for Image Enhancement

May 02, 2026

This work addresses the lack of a unified theoretical framework in existing image enhancement methods, which has led to ambiguous relationships and hindered fair comparisons. The authors propose the first unified perspective grounded in continuous-time stochastic differential equations (SDEs), systematically categorizing prevailing approaches into three classes: unconditional diffusion models, Ornstein–Uhlenbeck processes, and diffusion bridges. This formulation clarifies their fundamental differences in drift terms, diffusion coefficients, terminal distributions, and boundary conditions. Through controlled experiments using a consistent architecture and training protocol, the study reveals that performance is primarily governed by specific design choices rather than the methodological category itself. To facilitate standardized research, the authors also release ItoVision, an open-source modular library enabling fair evaluation and rapid prototyping.

0 citationsRead paper

A Probabilistic Consensus-Driven Approach for Robust Counterfactual Explanations

Apr 19, 2026

This work addresses the fragility of existing counterfactual explanation methods, which often fail under minor model perturbations and are typically restricted to specific model architectures, rely on costly hyperparameter tuning, or lack explicit robustness guarantees. To overcome these limitations, the authors propose a novel approach that jointly models the data distribution and decision boundaries through an ensemble of models, training a conditional normalizing flow using probabilistic consensus among the ensemble members. This framework introduces a tunable robustness parameter that flexibly controls the required proportion of ensemble agreement for the target class, without necessitating retraining. Experimental results demonstrate that the method significantly enhances the empirical robustness of generated counterfactuals under model perturbations while preserving high explanation quality.

0 citationsRead paper

Towards plausibility in time series counterfactual explanations

Mar 09, 2026

This work addresses the lack of temporal structural plausibility in counterfactual explanations for time series classification by proposing a novel input-space gradient-based optimization method. It introduces soft dynamic time warping (soft-DTW) as a constraint to enforce temporal coherence—a first in this context—and incorporates alignment with k-nearest neighbors from the target class. A multi-objective loss function jointly optimizes the counterfactual examples for validity, sparsity, proximity, and temporal structural fidelity. Experimental results demonstrate that the generated counterfactuals not only achieve high classification validity but also significantly outperform existing methods in aligning with the target class distribution and preserving realistic temporal structures.

0 citationsRead paper

Recent publications

Latest Papers

Mind the Gap: Quantifying the Domain Gap in Cross-Sensor Diffusion Super-Resolution

Jun 26, 2026

This study addresses the domain gap between synthetic training data and real-world remote sensing imagery in cross-sensor super-resolution. It presents the first systematic evaluation of diffusion models on real Sentinel-2 and PlanetScope images, uncovering critical challenges in synthetic-to-real domain transfer. Leveraging a large-scale geometrically and temporally aligned dataset, the work integrates five state-of-the-art diffusion architectures with self-supervised learning and introduces LPIPS-Sat, a novel perceptual metric tailored for quantifying the domain discrepancy in satellite imagery. The findings reveal that models trained on synthetic data suffer significant performance degradation when applied to real scenes, while those trained directly on real data encounter optimization difficulties and struggle to generalize across the physical and radiometric diversity inherent in actual satellite observations, thereby highlighting fundamental limitations of current approaches.

0 citationsRead paper

V4FinBench: Benchmarking Tabular Foundation Models, LLMs, and Standard Methods on Corporate Bankruptcy Prediction

May 11, 2026

This study addresses the limitations of existing public datasets for corporate bankruptcy prediction—namely, their small scale, extreme class imbalance, and lack of a unified multi-horizon evaluation benchmark—by introducing V4FinBench, the first large-scale, publicly available, cross-regional (covering four countries in China and Europe) benchmark. V4FinBench incorporates 131 features, six prediction horizons, and a composite default definition. Using this benchmark, the authors systematically evaluate tabular foundation models (e.g., TabPFN), Llama-3-8B fine-tuned with QLoRA, and traditional gradient boosting methods under imbalance-aware training protocols. Results show that domain-adapted TabPFN matches or exceeds gradient boosting models in F1 and ROC-AUC for long-term predictions and demonstrates strong cross-regional generalization, whereas Llama-3-8B exhibits consistently weaker performance, particularly over longer horizons.

0 citationsRead paper

Unifying Deep Stochastic Processes for Image Enhancement

May 02, 2026

This work addresses the lack of a unified theoretical framework in existing image enhancement methods, which has led to ambiguous relationships and hindered fair comparisons. The authors propose the first unified perspective grounded in continuous-time stochastic differential equations (SDEs), systematically categorizing prevailing approaches into three classes: unconditional diffusion models, Ornstein–Uhlenbeck processes, and diffusion bridges. This formulation clarifies their fundamental differences in drift terms, diffusion coefficients, terminal distributions, and boundary conditions. Through controlled experiments using a consistent architecture and training protocol, the study reveals that performance is primarily governed by specific design choices rather than the methodological category itself. To facilitate standardized research, the authors also release ItoVision, an open-source modular library enabling fair evaluation and rapid prototyping.

0 citationsRead paper

A Probabilistic Consensus-Driven Approach for Robust Counterfactual Explanations

Apr 19, 2026

This work addresses the fragility of existing counterfactual explanation methods, which often fail under minor model perturbations and are typically restricted to specific model architectures, rely on costly hyperparameter tuning, or lack explicit robustness guarantees. To overcome these limitations, the authors propose a novel approach that jointly models the data distribution and decision boundaries through an ensemble of models, training a conditional normalizing flow using probabilistic consensus among the ensemble members. This framework introduces a tunable robustness parameter that flexibly controls the required proportion of ensemble agreement for the target class, without necessitating retraining. Experimental results demonstrate that the method significantly enhances the empirical robustness of generated counterfactuals under model perturbations while preserving high explanation quality.

0 citationsRead paper

Towards plausibility in time series counterfactual explanations

Mar 09, 2026

This work addresses the lack of temporal structural plausibility in counterfactual explanations for time series classification by proposing a novel input-space gradient-based optimization method. It introduces soft dynamic time warping (soft-DTW) as a constraint to enforce temporal coherence—a first in this context—and incorporates alignment with k-nearest neighbors from the target class. A multi-objective loss function jointly optimizes the counterfactual examples for validity, sparsity, proximity, and temporal structural fidelity. Experimental results demonstrate that the generated counterfactuals not only achieve high classification validity but also significantly outperform existing methods in aligning with the target class distribution and preserving realistic temporal structures.

0 citationsRead paper