Selected work

Representative Papers

ANVIL: Accelerator-Native Video Interpolation via Codec Motion Vector Priors

Mar 27, 2026

This work addresses the challenges of deploying optical flow–based video frame interpolation on mobile devices, where incompatibility of sampling operators, iterative refinement failure under 8-bit quantization, and memory constraints hinder practicality. The authors propose a novel architecture that eliminates learned optical flow estimation by repurposing motion vectors from an H.264 decoder to pre-align input frames. By discarding spatial sampling and iterative accumulation, the method relies solely on a convolutional network to refine residual details, yielding a fully compute-intensive inference graph. This approach pioneers the use of codec-derived motion vectors as priors for real-time mobile interpolation, enabling a low-latency design fully compatible with neural processing units (NPUs) and uncovering the causal mechanism behind iterative optical flow failure under quantization. On a Snapdragon 8 Gen 3 platform, it achieves 12.8 ms latency for 1080p interpolation, with an end-to-end median latency of 28.4 ms in an open-source Android player, validated over 54,623 samples across 30 minutes of continuous playback.

0 citationsRead paper

MDD-Thinker: Towards Large Reasoning Models for Major Depressive Disorder Diagnosis

Sep 28, 2025

Depression diagnosis faces key challenges: heavy reliance on subjective clinical assessment, limited capability in integrating multimodal data (e.g., clinical notes, rating scales, biomarkers), poor model interpretability, and susceptibility to hallucination. To address these bottlenecks, we propose the first reasoning-augmented large language model (LLM) framework trained on large-scale real-world clinical data—including UK Biobank—comprising 50,000 inference samples. Our method synergistically combines supervised fine-tuning with human feedback-based reinforcement learning to enable coherent, multimodal clinical reasoning. On the binary MDD classification task, it achieves 0.827 accuracy and 0.808 F1-score—surpassing conventional models by over 29% relatively—while matching or exceeding the performance of significantly larger LMs. The framework delivers high accuracy, strong interpretability, and computational efficiency, demonstrating substantial potential for clinical deployment.

0 citationsRead paper

Using KL-Divergence to Focus Frequency Information in Low-Light Image Enhancement

Sep 16, 2025

To address insufficient frequency-domain modeling and global structural distortions caused by pixel-wise losses in low-light image enhancement, this paper proposes a distribution-aware Fourier-domain optimization framework. Methodologically: (1) a closed-form Fourier-domain loss based on KL divergence is designed to directly align the statistical distributions of amplitude and phase spectra between enhanced and ground-truth images; (2) a U-shaped network integrating cross-attention and gating mechanisms is introduced to jointly model spatial-frequency dependencies; (3) KL divergence is embedded into the VGG perceptual loss to enhance deep semantic structure fidelity. Extensive experiments demonstrate state-of-the-art performance across multiple benchmark datasets, with both quantitative metrics and qualitative visual results significantly surpassing existing methods. The source code is publicly available.

0 citationsRead paper

MDD-LLM: Towards Accuracy Large Language Models for Major Depressive Disorder Diagnosis

Apr 28, 2025

Globally, over 300 million individuals suffer from major depressive disorder (MDD), yet disparities in healthcare access and diagnostic complexity hinder timely, accurate identification—necessitating scalable, high-accuracy AI-assisted tools. This study pioneers the systematic application of large language models (LLMs) to clinical MDD risk prediction. Leveraging phenotypic data from 274,000 participants in the UK Biobank, we introduce a structured text encoding paradigm tailored for medical tabular data. A 70-billion-parameter LLM is supervised fine-tuned on this encoded representation and rigorously evaluated using accuracy, area under the ROC curve (AUC), and 95% confidence intervals. The resulting MDD-LLM achieves an accuracy of 0.8378 and an AUC of 0.8919 (95% CI: 0.8799–0.9040), significantly outperforming conventional machine learning and deep learning baselines. This work establishes a novel, empirically validated paradigm for deploying LLMs in precision psychiatry.

0 citationsRead paper

Recent publications

Latest Papers

ANVIL: Accelerator-Native Video Interpolation via Codec Motion Vector Priors

Mar 27, 2026

This work addresses the challenges of deploying optical flow–based video frame interpolation on mobile devices, where incompatibility of sampling operators, iterative refinement failure under 8-bit quantization, and memory constraints hinder practicality. The authors propose a novel architecture that eliminates learned optical flow estimation by repurposing motion vectors from an H.264 decoder to pre-align input frames. By discarding spatial sampling and iterative accumulation, the method relies solely on a convolutional network to refine residual details, yielding a fully compute-intensive inference graph. This approach pioneers the use of codec-derived motion vectors as priors for real-time mobile interpolation, enabling a low-latency design fully compatible with neural processing units (NPUs) and uncovering the causal mechanism behind iterative optical flow failure under quantization. On a Snapdragon 8 Gen 3 platform, it achieves 12.8 ms latency for 1080p interpolation, with an end-to-end median latency of 28.4 ms in an open-source Android player, validated over 54,623 samples across 30 minutes of continuous playback.

0 citationsRead paper

MDD-Thinker: Towards Large Reasoning Models for Major Depressive Disorder Diagnosis

Sep 28, 2025

Depression diagnosis faces key challenges: heavy reliance on subjective clinical assessment, limited capability in integrating multimodal data (e.g., clinical notes, rating scales, biomarkers), poor model interpretability, and susceptibility to hallucination. To address these bottlenecks, we propose the first reasoning-augmented large language model (LLM) framework trained on large-scale real-world clinical data—including UK Biobank—comprising 50,000 inference samples. Our method synergistically combines supervised fine-tuning with human feedback-based reinforcement learning to enable coherent, multimodal clinical reasoning. On the binary MDD classification task, it achieves 0.827 accuracy and 0.808 F1-score—surpassing conventional models by over 29% relatively—while matching or exceeding the performance of significantly larger LMs. The framework delivers high accuracy, strong interpretability, and computational efficiency, demonstrating substantial potential for clinical deployment.

0 citationsRead paper

Using KL-Divergence to Focus Frequency Information in Low-Light Image Enhancement

Sep 16, 2025

To address insufficient frequency-domain modeling and global structural distortions caused by pixel-wise losses in low-light image enhancement, this paper proposes a distribution-aware Fourier-domain optimization framework. Methodologically: (1) a closed-form Fourier-domain loss based on KL divergence is designed to directly align the statistical distributions of amplitude and phase spectra between enhanced and ground-truth images; (2) a U-shaped network integrating cross-attention and gating mechanisms is introduced to jointly model spatial-frequency dependencies; (3) KL divergence is embedded into the VGG perceptual loss to enhance deep semantic structure fidelity. Extensive experiments demonstrate state-of-the-art performance across multiple benchmark datasets, with both quantitative metrics and qualitative visual results significantly surpassing existing methods. The source code is publicly available.

0 citationsRead paper

MDD-LLM: Towards Accuracy Large Language Models for Major Depressive Disorder Diagnosis

Apr 28, 2025

Globally, over 300 million individuals suffer from major depressive disorder (MDD), yet disparities in healthcare access and diagnostic complexity hinder timely, accurate identification—necessitating scalable, high-accuracy AI-assisted tools. This study pioneers the systematic application of large language models (LLMs) to clinical MDD risk prediction. Leveraging phenotypic data from 274,000 participants in the UK Biobank, we introduce a structured text encoding paradigm tailored for medical tabular data. A 70-billion-parameter LLM is supervised fine-tuned on this encoded representation and rigorously evaluated using accuracy, area under the ROC curve (AUC), and 95% confidence intervals. The resulting MDD-LLM achieves an accuracy of 0.8378 and an AUC of 0.8919 (95% CI: 0.8799–0.9040), significantly outperforming conventional machine learning and deep learning baselines. This work establishes a novel, empirically validated paradigm for deploying LLMs in precision psychiatry.

0 citationsRead paper