Selected work

Representative Papers

Quality Adaptive Angular Margin Learning for Respiratory Sound Classification

Jun 10, 2026

This work addresses the challenge of limited feature generalizability in respiratory sound classification caused by variations in recording quality and class imbalance. To this end, the authors propose QLung, a novel framework that introduces, for the first time, a no-reference audio quality assessment metric based on spectral entropy and root-mean-square energy. This metric dynamically adjusts the angular margin in a normalized angular classifier and is combined with a log-scaling strategy to enhance intra-class compactness and inter-class separability, thereby stabilizing training. Evaluated on the ICBHI dataset, QLung achieves a 2.46% improvement over the cross-entropy baseline and demonstrates state-of-the-art out-of-distribution generalization performance on the SPRSound dataset.

0 citationsRead paper

2.5-D Decomposition for LLM-Based Spatial Construction

May 07, 2026

This work addresses the systematic spatial reasoning errors exhibited by large language models when generating 3D structures from natural language instructions, which often manifest as coordinate inaccuracies that undermine structural reliability. To mitigate this, the authors propose a neuro-symbolic 2.5-D decomposition approach that disentangles deterministic physical constraints—such as gravity—from the language model’s output. The model is restricted to planning layouts in a 2D plane, while a symbolic executor determines vertical stacking based on column occupancy. This strategy significantly improves construction accuracy, achieving a 94.6% average structural correctness on the Build What I Mean benchmark—surpassing GPT-4o (90.3%) and the previous state-of-the-art system (76.3%). Notably, it retains 94.5% performance on Jetson Thor AGX edge hardware. Ablation studies attribute a 50.7-percentage-point accuracy gain to the proposed method, highlighting its potential for generalization to other physically constrained assembly tasks.

0 citationsRead paper

Meta-Ensemble Learning with Diverse Data Splits for Improved Respiratory Sound Classification

Apr 27, 2026

This study addresses the limitations of existing respiratory sound classification models, which suffer from small-scale, low-diversity datasets and insufficient generalization due to high prediction correlation among base models trained on overlapping data. To overcome these challenges, the authors propose a meta-ensemble learning framework that enhances model diversity by training base models under distinct data partitions—fixed split versus five-fold cross-validation—and at different granularities—patient-level versus sample-level. A learnable meta-model is then introduced to fuse the outputs of these diverse base models. The proposed approach achieves a state-of-the-art score of 66.49% on the ICBHI benchmark and demonstrates superior out-of-distribution generalization on two external datasets, highlighting its potential for real-world clinical deployment.

0 citationsRead paper

Improving Respiratory Sound Classification with Architecture-Agnostic Knowledge Distillation from Ensembles

May 28, 2025

Small-scale and low-quality breath sound datasets hinder classification performance, while ensemble models—though effective—impose substantial inference overhead. To address this, we propose an architecture-agnostic soft-label knowledge distillation framework that efficiently transfers knowledge from multiple teachers to a lightweight student model. Our key contributions are threefold: (1) We empirically demonstrate for the first time that a single teacher—architecturally identical to the student—suffices to yield significant performance gains; (2) we show that only a few teachers are needed to achieve near-optimal improvement, thereby relaxing the conventional requirement of heterogeneous, high-capacity teachers; and (3) we integrate soft-label distillation, ensemble learning, and time-frequency feature modeling tailored to respiratory acoustics. Evaluated on the ICHBI dataset, our method achieves a new state-of-the-art score of 64.39 (+0.85), with an average architecture-wise performance gain exceeding 1.16.

0 citationsRead paper

Language-Agnostic Suicidal Risk Detection Using Large Language Models

May 26, 2025

Existing approaches to adolescent suicide risk identification suffer from strong language dependency and poor generalizability across languages. To address this, we propose the first language-agnostic cross-lingual risk assessment framework. Our method first standardizes spoken input via automatic speech recognition (ASR) into Chinese text, then leverages large language models to extract bilingual (Chinese–English) risk features; subsequently, language-specific pre-trained models are fine-tuned independently. Crucially, the framework requires no multilingual annotated data—enabling, for the first time, cross-lingual suicide risk detection with joint bilingual feature modeling and knowledge transfer. Experiments demonstrate that our approach matches monolingual baselines in accuracy while substantially improving cross-lingual robustness and generalization. It establishes a scalable, annotation-efficient paradigm for mental health screening in low-resource language settings.

0 citationsRead paper

Recent publications

Latest Papers

Quality Adaptive Angular Margin Learning for Respiratory Sound Classification

Jun 10, 2026

This work addresses the challenge of limited feature generalizability in respiratory sound classification caused by variations in recording quality and class imbalance. To this end, the authors propose QLung, a novel framework that introduces, for the first time, a no-reference audio quality assessment metric based on spectral entropy and root-mean-square energy. This metric dynamically adjusts the angular margin in a normalized angular classifier and is combined with a log-scaling strategy to enhance intra-class compactness and inter-class separability, thereby stabilizing training. Evaluated on the ICBHI dataset, QLung achieves a 2.46% improvement over the cross-entropy baseline and demonstrates state-of-the-art out-of-distribution generalization performance on the SPRSound dataset.

0 citationsRead paper

2.5-D Decomposition for LLM-Based Spatial Construction

May 07, 2026

This work addresses the systematic spatial reasoning errors exhibited by large language models when generating 3D structures from natural language instructions, which often manifest as coordinate inaccuracies that undermine structural reliability. To mitigate this, the authors propose a neuro-symbolic 2.5-D decomposition approach that disentangles deterministic physical constraints—such as gravity—from the language model’s output. The model is restricted to planning layouts in a 2D plane, while a symbolic executor determines vertical stacking based on column occupancy. This strategy significantly improves construction accuracy, achieving a 94.6% average structural correctness on the Build What I Mean benchmark—surpassing GPT-4o (90.3%) and the previous state-of-the-art system (76.3%). Notably, it retains 94.5% performance on Jetson Thor AGX edge hardware. Ablation studies attribute a 50.7-percentage-point accuracy gain to the proposed method, highlighting its potential for generalization to other physically constrained assembly tasks.

0 citationsRead paper

Meta-Ensemble Learning with Diverse Data Splits for Improved Respiratory Sound Classification

Apr 27, 2026

This study addresses the limitations of existing respiratory sound classification models, which suffer from small-scale, low-diversity datasets and insufficient generalization due to high prediction correlation among base models trained on overlapping data. To overcome these challenges, the authors propose a meta-ensemble learning framework that enhances model diversity by training base models under distinct data partitions—fixed split versus five-fold cross-validation—and at different granularities—patient-level versus sample-level. A learnable meta-model is then introduced to fuse the outputs of these diverse base models. The proposed approach achieves a state-of-the-art score of 66.49% on the ICBHI benchmark and demonstrates superior out-of-distribution generalization on two external datasets, highlighting its potential for real-world clinical deployment.

0 citationsRead paper

Improving Respiratory Sound Classification with Architecture-Agnostic Knowledge Distillation from Ensembles

May 28, 2025

Small-scale and low-quality breath sound datasets hinder classification performance, while ensemble models—though effective—impose substantial inference overhead. To address this, we propose an architecture-agnostic soft-label knowledge distillation framework that efficiently transfers knowledge from multiple teachers to a lightweight student model. Our key contributions are threefold: (1) We empirically demonstrate for the first time that a single teacher—architecturally identical to the student—suffices to yield significant performance gains; (2) we show that only a few teachers are needed to achieve near-optimal improvement, thereby relaxing the conventional requirement of heterogeneous, high-capacity teachers; and (3) we integrate soft-label distillation, ensemble learning, and time-frequency feature modeling tailored to respiratory acoustics. Evaluated on the ICHBI dataset, our method achieves a new state-of-the-art score of 64.39 (+0.85), with an average architecture-wise performance gain exceeding 1.16.

0 citationsRead paper

Language-Agnostic Suicidal Risk Detection Using Large Language Models

May 26, 2025

Existing approaches to adolescent suicide risk identification suffer from strong language dependency and poor generalizability across languages. To address this, we propose the first language-agnostic cross-lingual risk assessment framework. Our method first standardizes spoken input via automatic speech recognition (ASR) into Chinese text, then leverages large language models to extract bilingual (Chinese–English) risk features; subsequently, language-specific pre-trained models are fine-tuned independently. Crucially, the framework requires no multilingual annotated data—enabling, for the first time, cross-lingual suicide risk detection with joint bilingual feature modeling and knowledge transfer. Experiments demonstrate that our approach matches monolingual baselines in accuracy while substantially improving cross-lingual robustness and generalization. It establishes a scalable, annotation-efficient paradigm for mental health screening in low-resource language settings.

0 citationsRead paper