Selected work

Representative Papers

1BT: One-Block Transformer for EEG-Based Cognitive Workload Assessment

Apr 21, 2026

This work addresses the challenge of achieving both high accuracy and computational efficiency in cognitive workload assessment under resource-constrained conditions. The authors propose an extremely lightweight single-block Transformer architecture (1BT), which, for the first time, applies a single Transformer block to multi-channel EEG time-series modeling. By incorporating a latent bottleneck to compress input signals and integrating lightweight self-attention and cross-attention mechanisms, the model enables efficient discriminative learning. Requiring only 0.5 million parameters and 0.02 GFLOPs, the proposed method attains competitive cognitive workload classification performance while drastically reducing model size and computational overhead, making it well-suited for real-time, low-power deployment scenarios.

2 citationsRead paper

Design principles to Increase Technology Self-efficacy for Older Australians with Mild Cognitive Impairment (MCI) and Older Carers

Aug 31, 2026

研究探讨了澳大利亚轻度认知障碍老年人及其照顾者采用和持续使用技术的动机和促进因素，提出基于自我效能理论的设计原则以提高技术自信心。

0 citationsRead paper

A Multi-Scale Temporal Framework with Dynamic Fusion for EEG-Based Emotion Recognition

Aug 09, 2026

This study addresses the limitations of existing EEG-based emotion recognition methods, which typically rely on a single time window and struggle to capture the multi-scale temporal dynamics inherent in complex emotional states such as mixed emotions. To overcome this, the authors propose a multi-scale temporal modeling framework that decomposes EEG signals into multiple time windows, extracts features using a shared attention-based encoder, and employs a dynamic fusion module to adaptively assign sample-specific weights across scales. The approach is the first to demonstrate the effectiveness of multi-scale modeling in a three-class emotion recognition task involving mixed emotions. Under a subject-independent protocol, the method achieves classification accuracies of 65.22% for binary classification and 45.43% for ternary classification, significantly outperforming baseline models using the full-length signal.

0 citationsRead paper

From Evaluation to Optimisation: Hierarchy-Aware Training Signals for CWE Prediction in Python

Jul 23, 2026

This study addresses the performance degradation of CWE-level vulnerability prediction in Python code under distribution shift. It presents the first empirical validation that the hierarchical evaluation penalty defined in the ALPHA benchmark can be effectively leveraged as a training signal. The authors systematically compare three training mechanisms: supervised fine-tuning, dual-head classification loss, and GRPO reinforcement learning incorporating normalized ALPHA penalties. Experimental results demonstrate that GRPO significantly outperforms supervised approaches under distribution shift. On the Qwen2.5-Coder-7B model, the best GRPO policy reduces cumulative ALPHA penalty by 27.9% with greedy decoding and 25.5% with sampling decoding on the SVEN dataset, achieving statistically equivalent performance to a zero-shot teacher model 4.5 times larger in scale.

0 citationsRead paper

ReFace: Reorganizing Facial Spatiotemporal Representations for Improved Pain Assessment

Jul 21, 2026

This study addresses the challenge of automatic pain assessment arising from the spatial heterogeneity of facial pain cues. To this end, the authors propose ReFace, a novel approach that partitions facial videos into four spatial quadrants and performs tokenization and spatiotemporal modeling on each quadrant independently, rather than processing the entire face as a whole. This design enables more precise capture of localized pain-related features. ReFace further introduces an innovative spatial recombination strategy that enhances model performance without increasing the total number of pixels. Notably, the method demonstrates that using only a single quadrant can maintain competitive accuracy while substantially reducing computational cost. Evaluated on the AI4Pain dataset under the standard benchmark protocol, ReFace achieves a test accuracy of 56.00% using video input alone, setting a new state-of-the-art result.

0 citationsRead paper

Recent publications

Latest Papers

Design principles to Increase Technology Self-efficacy for Older Australians with Mild Cognitive Impairment (MCI) and Older Carers

Aug 31, 2026

研究探讨了澳大利亚轻度认知障碍老年人及其照顾者采用和持续使用技术的动机和促进因素，提出基于自我效能理论的设计原则以提高技术自信心。

0 citationsRead paper

A Multi-Scale Temporal Framework with Dynamic Fusion for EEG-Based Emotion Recognition

Aug 09, 2026

This study addresses the limitations of existing EEG-based emotion recognition methods, which typically rely on a single time window and struggle to capture the multi-scale temporal dynamics inherent in complex emotional states such as mixed emotions. To overcome this, the authors propose a multi-scale temporal modeling framework that decomposes EEG signals into multiple time windows, extracts features using a shared attention-based encoder, and employs a dynamic fusion module to adaptively assign sample-specific weights across scales. The approach is the first to demonstrate the effectiveness of multi-scale modeling in a three-class emotion recognition task involving mixed emotions. Under a subject-independent protocol, the method achieves classification accuracies of 65.22% for binary classification and 45.43% for ternary classification, significantly outperforming baseline models using the full-length signal.

0 citationsRead paper

From Evaluation to Optimisation: Hierarchy-Aware Training Signals for CWE Prediction in Python

Jul 23, 2026

This study addresses the performance degradation of CWE-level vulnerability prediction in Python code under distribution shift. It presents the first empirical validation that the hierarchical evaluation penalty defined in the ALPHA benchmark can be effectively leveraged as a training signal. The authors systematically compare three training mechanisms: supervised fine-tuning, dual-head classification loss, and GRPO reinforcement learning incorporating normalized ALPHA penalties. Experimental results demonstrate that GRPO significantly outperforms supervised approaches under distribution shift. On the Qwen2.5-Coder-7B model, the best GRPO policy reduces cumulative ALPHA penalty by 27.9% with greedy decoding and 25.5% with sampling decoding on the SVEN dataset, achieving statistically equivalent performance to a zero-shot teacher model 4.5 times larger in scale.

0 citationsRead paper

ReFace: Reorganizing Facial Spatiotemporal Representations for Improved Pain Assessment

Jul 21, 2026

This study addresses the challenge of automatic pain assessment arising from the spatial heterogeneity of facial pain cues. To this end, the authors propose ReFace, a novel approach that partitions facial videos into four spatial quadrants and performs tokenization and spatiotemporal modeling on each quadrant independently, rather than processing the entire face as a whole. This design enables more precise capture of localized pain-related features. ReFace further introduces an innovative spatial recombination strategy that enhances model performance without increasing the total number of pixels. Notably, the method demonstrates that using only a single quadrant can maintain competitive accuracy while substantially reducing computational cost. Evaluated on the AI4Pain dataset under the standard benchmark protocol, ReFace achieves a test accuracy of 56.00% using video input alone, setting a new state-of-the-art result.

0 citationsRead paper

1BT: One-Block Transformer for EEG-Based Cognitive Workload Assessment

Apr 21, 2026

This work addresses the challenge of achieving both high accuracy and computational efficiency in cognitive workload assessment under resource-constrained conditions. The authors propose an extremely lightweight single-block Transformer architecture (1BT), which, for the first time, applies a single Transformer block to multi-channel EEG time-series modeling. By incorporating a latent bottleneck to compress input signals and integrating lightweight self-attention and cross-attention mechanisms, the model enables efficient discriminative learning. Requiring only 0.5 million parameters and 0.02 GFLOPs, the proposed method attains competitive cognitive workload classification performance while drastically reducing model size and computational overhead, making it well-suited for real-time, low-power deployment scenarios.

2 citationsRead paper