Selected work

Representative Papers

ODE-Based Transformer Decoders for Iterative Sign Language Translation

Aug 11, 2026

This work addresses the high computational cost and low parameter efficiency of current sign language translation models, which often rely on scaling up model size for performance gains. The authors propose a novel reconstruction of the Transformer decoder from the perspective of ordinary differential equations (ODEs), introducing higher-order numerical integration schemes—such as Runge-Kutta methods RK-2 and RK-4—into sign language translation for the first time. This approach replaces conventional residual connections with more accurate and stable iterative optimization, enhancing model expressiveness without increasing parameter count. Experimental results demonstrate that the proposed method achieves BLEU-4 scores of 22.96 and 19.34 on the PHOENIX-2014-T and CSL-Daily datasets, respectively, outperforming baseline models while using fewer decoder layers and iteration steps.

0 citationsRead paper

SCR-Guided Difficulty-Aware Optimization for Infrared Small Target Detection

Jun 17, 2026

This work addresses the challenges of infrared small target detection, which include strong background clutter, low contrast, and weak spatial response, rendering conventional geometry-based overlap metrics inadequate for accurate performance evaluation. To overcome this limitation, the authors propose the REEM framework, which—by introducing signal-to-clutter ratio (SCR) as a physically grounded visibility prior into the training process—modulates the soft-IoU learning signal in a difficulty-aware manner via differentiable SCR computation. This approach enhances the model’s focus on low-visibility targets without altering network architecture or increasing inference overhead. Evaluated on the U-Net-based MSHNet, the method demonstrates significant improvements in detection probability and IoU while substantially reducing false alarm rates, with particularly pronounced gains in low-SCR scenarios.

0 citationsRead paper

Yolo-Key-6D: Single Stage Monocular 6D Pose Estimation with Keypoint Enhancements

Mar 04, 2026

This work proposes a single-stage, end-to-end framework based on YOLO to address the high latency of multi-stage methods for 6D object pose estimation from monocular RGB images. The approach introduces an auxiliary keypoint head to regress 2D projections of 3D bounding box corners and employs a continuous 9D representation combined with singular value decomposition (SVD) to enable stable and differentiable rotation regression. By integrating a keypoint enhancement mechanism and the 9D→SO(3) rotation representation into a single-stage detector, the method achieves a favorable balance between accuracy and efficiency. Evaluated on the LINEMOD and LINEMOD-Occluded datasets, it attains ADD(-S) 0.1d accuracies of 96.24% and 69.41%, respectively, while meeting real-time performance requirements.

0 citationsRead paper

FewMMBench: A Benchmark for Multimodal Few-Shot Learning

Feb 25, 2026

This work addresses the lack of systematic evaluation of multimodal large language models (MLLMs) in few-shot learning scenarios, particularly concerning in-context learning and chain-of-thought prompting. To this end, we introduce FewMMBench—the first comprehensive benchmark dedicated to multimodal few-shot learning—encompassing diverse tasks such as attribute recognition and temporal reasoning. We conduct a systematic evaluation of 26 open-source models across six model families under zero-shot, few-shot, and chain-of-thought enhanced settings. Our findings reveal that instruction-tuned models, despite strong zero-shot capabilities, exhibit limited sensitivity—or even performance degradation—when provided with examples or chain-of-thought prompts. Moreover, gains from retrieval-based example selection and increased context length are marginal, underscoring the value of FewMMBench as a diagnostic tool for multimodal few-shot learning.

0 citationsRead paper

A Deep Reinforcement Learning-Based TCP Congestion Control Algorithm: Design, Simulation, and Evaluation

Aug 01, 2025

Traditional TCP congestion control mechanisms (e.g., NewReno) suffer from poor adaptability, high latency, and low throughput in dynamic network environments. To address these limitations, this paper proposes an end-to-end, deep reinforcement learning–based approach for adaptive congestion window (cWnd) control. It introduces the Deep Q-Network (DQN) into TCP congestion control for the first time, leveraging a co-simulation framework integrating NS-3 and OpenGym to enable real-time online perception and decision-making based on dynamic network states—including round-trip delay, packet loss, and bandwidth fluctuations. Departing from fixed heuristic rules, the method employs a weighted reward function combining throughput and latency to guide policy learning, significantly enhancing generalization and responsiveness across heterogeneous network scenarios. Experimental results demonstrate that, compared to NewReno, the proposed scheme reduces average end-to-end latency by 32%, improves throughput by approximately 28%, and exhibits superior stability.

0 citationsRead paper

Recent publications

Latest Papers

ODE-Based Transformer Decoders for Iterative Sign Language Translation

Aug 11, 2026

This work addresses the high computational cost and low parameter efficiency of current sign language translation models, which often rely on scaling up model size for performance gains. The authors propose a novel reconstruction of the Transformer decoder from the perspective of ordinary differential equations (ODEs), introducing higher-order numerical integration schemes—such as Runge-Kutta methods RK-2 and RK-4—into sign language translation for the first time. This approach replaces conventional residual connections with more accurate and stable iterative optimization, enhancing model expressiveness without increasing parameter count. Experimental results demonstrate that the proposed method achieves BLEU-4 scores of 22.96 and 19.34 on the PHOENIX-2014-T and CSL-Daily datasets, respectively, outperforming baseline models while using fewer decoder layers and iteration steps.

0 citationsRead paper

SCR-Guided Difficulty-Aware Optimization for Infrared Small Target Detection

Jun 17, 2026

This work addresses the challenges of infrared small target detection, which include strong background clutter, low contrast, and weak spatial response, rendering conventional geometry-based overlap metrics inadequate for accurate performance evaluation. To overcome this limitation, the authors propose the REEM framework, which—by introducing signal-to-clutter ratio (SCR) as a physically grounded visibility prior into the training process—modulates the soft-IoU learning signal in a difficulty-aware manner via differentiable SCR computation. This approach enhances the model’s focus on low-visibility targets without altering network architecture or increasing inference overhead. Evaluated on the U-Net-based MSHNet, the method demonstrates significant improvements in detection probability and IoU while substantially reducing false alarm rates, with particularly pronounced gains in low-SCR scenarios.

0 citationsRead paper

Yolo-Key-6D: Single Stage Monocular 6D Pose Estimation with Keypoint Enhancements

Mar 04, 2026

This work proposes a single-stage, end-to-end framework based on YOLO to address the high latency of multi-stage methods for 6D object pose estimation from monocular RGB images. The approach introduces an auxiliary keypoint head to regress 2D projections of 3D bounding box corners and employs a continuous 9D representation combined with singular value decomposition (SVD) to enable stable and differentiable rotation regression. By integrating a keypoint enhancement mechanism and the 9D→SO(3) rotation representation into a single-stage detector, the method achieves a favorable balance between accuracy and efficiency. Evaluated on the LINEMOD and LINEMOD-Occluded datasets, it attains ADD(-S) 0.1d accuracies of 96.24% and 69.41%, respectively, while meeting real-time performance requirements.

0 citationsRead paper

FewMMBench: A Benchmark for Multimodal Few-Shot Learning

Feb 25, 2026

This work addresses the lack of systematic evaluation of multimodal large language models (MLLMs) in few-shot learning scenarios, particularly concerning in-context learning and chain-of-thought prompting. To this end, we introduce FewMMBench—the first comprehensive benchmark dedicated to multimodal few-shot learning—encompassing diverse tasks such as attribute recognition and temporal reasoning. We conduct a systematic evaluation of 26 open-source models across six model families under zero-shot, few-shot, and chain-of-thought enhanced settings. Our findings reveal that instruction-tuned models, despite strong zero-shot capabilities, exhibit limited sensitivity—or even performance degradation—when provided with examples or chain-of-thought prompts. Moreover, gains from retrieval-based example selection and increased context length are marginal, underscoring the value of FewMMBench as a diagnostic tool for multimodal few-shot learning.

0 citationsRead paper

A Deep Reinforcement Learning-Based TCP Congestion Control Algorithm: Design, Simulation, and Evaluation

Aug 01, 2025

Traditional TCP congestion control mechanisms (e.g., NewReno) suffer from poor adaptability, high latency, and low throughput in dynamic network environments. To address these limitations, this paper proposes an end-to-end, deep reinforcement learning–based approach for adaptive congestion window (cWnd) control. It introduces the Deep Q-Network (DQN) into TCP congestion control for the first time, leveraging a co-simulation framework integrating NS-3 and OpenGym to enable real-time online perception and decision-making based on dynamic network states—including round-trip delay, packet loss, and bandwidth fluctuations. Departing from fixed heuristic rules, the method employs a weighted reward function combining throughput and latency to guide policy learning, significantly enhancing generalization and responsiveness across heterogeneous network scenarios. Experimental results demonstrate that, compared to NewReno, the proposed scheme reduces average end-to-end latency by 32%, improves throughput by approximately 28%, and exhibits superior stability.

0 citationsRead paper