ARNAI: Artifact Removal Network based on Autoencoding and Inpainting for Robust Spinal Image Segmentation and Measurement
本文提出了一种基于自编码和修复的ARNAI网络，用于去除脊柱影像中的植入物伪影，提高了术后脊柱参数测量的准确性。
0 citationsRead paper
Selected work
Representative Papers
Missingness as Signal: Channel-Independent Spectrogram Learning for Clinical Time Series Prediction
This study addresses the challenges of clinical multivariate time series prediction, where heterogeneous physiological measurements and missing data are prevalent, and conventional approaches often overlook the informative patterns embedded in missingness that reflect clinical decision-making and disease severity. To this end, the authors propose the CISM framework, which uniquely treats missingness patterns as structured signals and aligns them with variable-level time-frequency spectrograms. The method explicitly integrates observed values and missingness streams through channel-independent time-frequency spectrogram generation, variable-aligned encoding, and pixel-level mask injection. Evaluated on the MIMIC-IV in-hospital mortality prediction task, CISM achieves an AUROC of 0.7225, AUPRC of 0.3308, and F1 score of 0.3808, significantly outperforming existing baselines and demonstrating the independent predictive value of missingness patterns.
0 citationsRead paper
A Precedent-Guided Co-Scientist for Side-Effect-Aware Drug Redesign
This study addresses the challenge of precisely mitigating specific adverse drug effects while preserving therapeutic efficacy by introducing PRECEDE, a novel framework that incorporates precedent-guided, auditable, and falsifiable reasoning into AI-assisted drug redesign for the first time. PRECEDE employs a large language model as an orchestrator, integrating drug–side effect association data, biomedical knowledge graphs, and historical precedents of safety optimization, while embedding human-in-the-loop review to ensure generated proposals remain within established pharmacological boundaries. Experimental results demonstrate that PRECEDE produces interpretable and traceable drug redesign strategies that effectively balance the retention of therapeutic activity with the alleviation of targeted side effects.
0 citationsRead paper
Multimodal Forecasting for Commodity Prices Using Spectrogram-Based and Time Series Representations
This work addresses the challenge of commodity price forecasting in multivariate time series, where complex cross-variable dependencies and interference from heterogeneous external factors hinder prediction accuracy. To tackle this, the authors propose a multimodal modeling paradigm that integrates time-frequency and temporal modalities. Specifically, Morlet wavelet transform is employed to generate spectrograms, from which frequency-aware features are extracted using a Vision Transformer. Exogenous variables are encoded via a Transformer module, and a bidirectional cross-attention mechanism is introduced to unify multimodal representations. This approach effectively captures cross-modal interactions while preserving modality-specific characteristics, significantly enhancing the model’s ability to discern multiscale dynamic patterns in financial time series. Extensive experiments demonstrate that the proposed method consistently outperforms seven state-of-the-art baselines across multiple commodity price prediction tasks, prediction horizons, and evaluation metrics.
0 citationsRead paper
Sentence Curve Language Models
This work proposes a novel approach to language modeling by representing sentences as continuous spline curves, where shared control points jointly govern multiple tokens to enable coherent modeling of both local details and global semantic structure. Unlike conventional and diffusion-based language models that rely on static word embeddings and struggle to capture sentence-level coherence, the proposed method integrates continuous sentence representations into a diffusion language modeling framework. This integration not only enhances structural awareness but also induces an implicit regularization effect, enabling stable training without knowledge distillation. Experimental results demonstrate state-of-the-art performance among diffusion language models on IWSLT14 and WMT14 benchmarks, and superior results over existing discrete diffusion approaches on LM1B.
0 citationsRead paper
Recent publications
Latest Papers
ARNAI: Artifact Removal Network based on Autoencoding and Inpainting for Robust Spinal Image Segmentation and Measurement
本文提出了一种基于自编码和修复的ARNAI网络，用于去除脊柱影像中的植入物伪影，提高了术后脊柱参数测量的准确性。
0 citationsRead paper
Missingness as Signal: Channel-Independent Spectrogram Learning for Clinical Time Series Prediction
This study addresses the challenges of clinical multivariate time series prediction, where heterogeneous physiological measurements and missing data are prevalent, and conventional approaches often overlook the informative patterns embedded in missingness that reflect clinical decision-making and disease severity. To this end, the authors propose the CISM framework, which uniquely treats missingness patterns as structured signals and aligns them with variable-level time-frequency spectrograms. The method explicitly integrates observed values and missingness streams through channel-independent time-frequency spectrogram generation, variable-aligned encoding, and pixel-level mask injection. Evaluated on the MIMIC-IV in-hospital mortality prediction task, CISM achieves an AUROC of 0.7225, AUPRC of 0.3308, and F1 score of 0.3808, significantly outperforming existing baselines and demonstrating the independent predictive value of missingness patterns.
0 citationsRead paper
A Precedent-Guided Co-Scientist for Side-Effect-Aware Drug Redesign
This study addresses the challenge of precisely mitigating specific adverse drug effects while preserving therapeutic efficacy by introducing PRECEDE, a novel framework that incorporates precedent-guided, auditable, and falsifiable reasoning into AI-assisted drug redesign for the first time. PRECEDE employs a large language model as an orchestrator, integrating drug–side effect association data, biomedical knowledge graphs, and historical precedents of safety optimization, while embedding human-in-the-loop review to ensure generated proposals remain within established pharmacological boundaries. Experimental results demonstrate that PRECEDE produces interpretable and traceable drug redesign strategies that effectively balance the retention of therapeutic activity with the alleviation of targeted side effects.
0 citationsRead paper
Multimodal Forecasting for Commodity Prices Using Spectrogram-Based and Time Series Representations
This work addresses the challenge of commodity price forecasting in multivariate time series, where complex cross-variable dependencies and interference from heterogeneous external factors hinder prediction accuracy. To tackle this, the authors propose a multimodal modeling paradigm that integrates time-frequency and temporal modalities. Specifically, Morlet wavelet transform is employed to generate spectrograms, from which frequency-aware features are extracted using a Vision Transformer. Exogenous variables are encoded via a Transformer module, and a bidirectional cross-attention mechanism is introduced to unify multimodal representations. This approach effectively captures cross-modal interactions while preserving modality-specific characteristics, significantly enhancing the model’s ability to discern multiscale dynamic patterns in financial time series. Extensive experiments demonstrate that the proposed method consistently outperforms seven state-of-the-art baselines across multiple commodity price prediction tasks, prediction horizons, and evaluation metrics.
0 citationsRead paper
Sentence Curve Language Models
This work proposes a novel approach to language modeling by representing sentences as continuous spline curves, where shared control points jointly govern multiple tokens to enable coherent modeling of both local details and global semantic structure. Unlike conventional and diffusion-based language models that rely on static word embeddings and struggle to capture sentence-level coherence, the proposed method integrates continuous sentence representations into a diffusion language modeling framework. This integration not only enhances structural awareness but also induces an implicit regularization effect, enabling stable training without knowledge distillation. Experimental results demonstrate state-of-the-art performance among diffusion language models on IWSLT14 and WMT14 benchmarks, and superior results over existing discrete diffusion approaches on LM1B.
0 citationsRead paper