Shenzhen Institute of Advanced Technology

Selected work

Representative Papers

SMAFormer: Synergistic Multi-Attention Transformer for Medical Image Segmentation

Aug 31, 2024IEEE International Conference on Bioinformatics and Biomedicine

To address the low segmentation accuracy of irregularly shaped and minute tumors in medical images, this paper proposes a Synergistic Multi-Attention Transformer (SMA-Transformer). The architecture innovatively integrates pixel-wise, channel-wise, and spatial-wise attention mechanisms, coupled with a feature fusion modulator to jointly model local details and global context—thereby mitigating information loss during attention transformation and feature recalibration. Its modular design incorporates residual connections to enhance gradient flow and feature reuse. Evaluated on multi-organ, liver tumor, and bladder tumor segmentation tasks, the method achieves state-of-the-art (SOTA) performance, particularly for small-target segmentation. It delivers significant improvements in key metrics: Dice coefficient and 95th-percentile Hausdorff Distance (HD95), demonstrating superior accuracy and boundary localization.

6 citationsRead paper

Seal2Real: Prompt Prior Learning on Diffusion Model for Unsupervised Document Seal Data Generation and Realisation

Oct 01, 2023arXiv.org

Document seal processing tasks—including segmentation, authenticity verification, removal, and occluded text recognition—are severely hindered by the scarcity of real-world annotated data. To address this, we propose the first end-to-end unsupervised seal image generation framework, built upon Stable Diffusion. Our method introduces a novel prompt-based prior learning mechanism that enables structural controllability and high-fidelity synthesis without requiring paired real seal samples. Leveraging this framework, we construct Seal-DB—the first large-scale, fully annotated seal dataset comprising 20,000 images. Extensive evaluation on Seal-DB demonstrates significant improvements in downstream tasks: seal segmentation and occluded text recognition accuracy increase by 12.6%–18.3%. Moreover, expert blind evaluation confirms that 91.4% of generated seals are perceived as photorealistic. This work establishes a foundational resource and methodology for data-starved seal analysis research.

4 citationsRead paper

From Completion to Editing: Unlocking Context-Aware Code Infilling via Search-and-Replace Instruction Tuning

Jan 19, 2026

Existing code completion methods, such as Fill-in-the-Middle (FIM), struggle to correct contextual errors and rely on potentially unsafe base models. Meanwhile, chat-based large language models suffer from performance degradation, and agent-based workflows incur high latency. To address these limitations, this work proposes the Search-and-Replace Infilling (SRI) framework, which extends code completion from static infilling to context-aware dynamic editing. SRI internalizes the agent-like verify-and-edit mechanism into a single inference pass, preserving low latency and general programming proficiency while maintaining instruction-following capabilities. Leveraging a synthetically constructed SRI-200K dataset and structured search-replace instructions, a model fine-tuned with only 20,000 samples—SRI-Coder—outperforms base models in completion accuracy while matching the inference speed of standard FIM.

2 citationsRead paper

MegaFlow: Large-Scale Distributed Orchestration System for the Agentic Era

Jan 12, 2026

This work addresses the current lack of open-source infrastructure capable of efficiently training and evaluating large-scale agents on complex tasks such as software engineering and computer operation. To this end, we propose a three-service decoupled architecture tailored for agent-environment interaction workloads, which separates the system into three independent services—model, agent, and environment—enabling fine-grained task scheduling, dynamic resource allocation, and unified interface communication. This design allows each component to scale independently and configure resources flexibly, significantly improving training efficiency and resource utilization. Experimental results demonstrate that the system can stably support tens of thousands of concurrent agent tasks, thereby filling a critical gap in infrastructure for large-scale agent training.

2 citationsRead paper

Visual Merit or Linguistic Crutch? A Close Look at DeepSeek-OCR

Jan 07, 2026arXiv.org

This study investigates whether DeepSeek-OCR relies on genuine visual capabilities or language priors under high-ratio visual-text compression and evaluates its reliability in long-context scenarios. By employing sentence- and word-level semantic perturbations to disentangle linguistic priors, combined with semantic corruption tests, context stress evaluations, and vision-language decoupling analyses, the work provides the first empirical evidence of the severe dependence of end-to-end OCR models on language priors: accuracy drops precipitously from 90% to 20% without linguistic support, hallucinations increase as visual tokens decrease, and the model completely fails at around 10,000 text tokens. In contrast, traditional OCR methods demonstrate greater robustness. The study also establishes a comprehensive multi-model robustness benchmark encompassing 13 baseline approaches.

1 citations1 influentialRead paper

Recent publications