Selected work

Representative Papers

Harnessing LLMs for Document-Guided Fuzzing of Python Libraries

Aug 12, 2026

This work addresses the challenge of testing Python library APIs, which often involve complex parameter constraints and cross-parameter dependencies that existing methods struggle to model effectively. The authors propose VistaFuzz, a novel approach that leverages locally deployed open-source large language models to automatically extract parameter specifications and dependency relationships directly from API documentation. These extracted constraints are then integrated into the fuzzing input generation process, enabling documentation-guided, high-efficiency testing. By combining natural language parsing, constraint reasoning, and fuzz testing, VistaFuzz achieves over 95% effective test case generation across 7,718 APIs from 12 widely used Python libraries. The evaluation uncovered 74 distinct issues, 43 of which were confirmed by developers and 29 subsequently fixed.

0 citationsRead paper

Spatially Grounded Concept Bottleneck Models for Trustworthy Breast Ultrasound Diagnosis

Jul 22, 2026

This study addresses the unreliability of explanations in conventional Concept Bottleneck Models (CBMs) for breast ultrasound diagnosis, which arises from misaligned concept activations due to the absence of spatial supervision. To resolve this, the authors propose the Spatially Guided Concept Bottleneck Model (SG-CBM), which introduces lesion masks partitioned into intra-lesion and posterior acoustic shadow regions. Leveraging coarse lesion contours as weak spatial supervision, SG-CBM employs grouped spatial anchoring objectives to guide clinically relevant concepts—such as morphology and posterior acoustic shadowing—to activate at anatomically plausible locations. A linear bottleneck classifier is integrated to jointly preserve semantic fidelity and spatial alignment. Experiments demonstrate that SG-CBM significantly improves both diagnostic AUROC and macro concept AUROC, while markedly enhancing the spatial alignment of concept evidence, thereby underscoring the critical role of high-quality spatial supervision in developing trustworthy medical AI systems.

0 citationsRead paper

Supervised Contrastive Frame Aggregation for Video Representation Learning

Dec 13, 2025

To address the high computational cost of temporal modeling and insufficient exploitation of global temporal context in video representation learning, this paper proposes the Supervised Contrastive Frame Aggregation (SCFA) framework. SCFA spatially arranges multiple frames into a single pseudo-image, enabling efficient reuse of lightweight image backbones (e.g., ResNet-50) and avoiding the prohibitive computational overhead of video Transformers. A supervised contrastive loss—guided by class labels—is introduced, and natural positive sample pairs are constructed via multi-temporal sampling within the same video to explicitly encode global temporal structure. The framework supports end-to-end video classification and captioning. On Penn Action and HMDB51, SCFA achieves 76% and 48% top-1 accuracy, respectively—substantially outperforming ViViT (43%/37%)—while significantly improving both training and inference efficiency.

0 citationsRead paper

Recent publications

Latest Papers

Harnessing LLMs for Document-Guided Fuzzing of Python Libraries

Aug 12, 2026

This work addresses the challenge of testing Python library APIs, which often involve complex parameter constraints and cross-parameter dependencies that existing methods struggle to model effectively. The authors propose VistaFuzz, a novel approach that leverages locally deployed open-source large language models to automatically extract parameter specifications and dependency relationships directly from API documentation. These extracted constraints are then integrated into the fuzzing input generation process, enabling documentation-guided, high-efficiency testing. By combining natural language parsing, constraint reasoning, and fuzz testing, VistaFuzz achieves over 95% effective test case generation across 7,718 APIs from 12 widely used Python libraries. The evaluation uncovered 74 distinct issues, 43 of which were confirmed by developers and 29 subsequently fixed.

0 citationsRead paper

Spatially Grounded Concept Bottleneck Models for Trustworthy Breast Ultrasound Diagnosis

Jul 22, 2026

This study addresses the unreliability of explanations in conventional Concept Bottleneck Models (CBMs) for breast ultrasound diagnosis, which arises from misaligned concept activations due to the absence of spatial supervision. To resolve this, the authors propose the Spatially Guided Concept Bottleneck Model (SG-CBM), which introduces lesion masks partitioned into intra-lesion and posterior acoustic shadow regions. Leveraging coarse lesion contours as weak spatial supervision, SG-CBM employs grouped spatial anchoring objectives to guide clinically relevant concepts—such as morphology and posterior acoustic shadowing—to activate at anatomically plausible locations. A linear bottleneck classifier is integrated to jointly preserve semantic fidelity and spatial alignment. Experiments demonstrate that SG-CBM significantly improves both diagnostic AUROC and macro concept AUROC, while markedly enhancing the spatial alignment of concept evidence, thereby underscoring the critical role of high-quality spatial supervision in developing trustworthy medical AI systems.

0 citationsRead paper

Supervised Contrastive Frame Aggregation for Video Representation Learning

Dec 13, 2025

To address the high computational cost of temporal modeling and insufficient exploitation of global temporal context in video representation learning, this paper proposes the Supervised Contrastive Frame Aggregation (SCFA) framework. SCFA spatially arranges multiple frames into a single pseudo-image, enabling efficient reuse of lightweight image backbones (e.g., ResNet-50) and avoiding the prohibitive computational overhead of video Transformers. A supervised contrastive loss—guided by class labels—is introduced, and natural positive sample pairs are constructed via multi-temporal sampling within the same video to explicitly encode global temporal structure. The framework supports end-to-end video classification and captioning. On Penn Action and HMDB51, SCFA achieves 76% and 48% top-1 accuracy, respectively—substantially outperforming ViViT (43%/37%)—while significantly improving both training and inference efficiency.

0 citationsRead paper