Projects

Some projects I have led or played a key role include:

		OpenDeRisk
	Ling-Coder-Lite

------------------------------------------

•   OpenDeRisk

  OpenDeRisk AI-Native Risk Intelligence Systems —— Your application system risk intelligent manager provides 7 * 24-hour comprehensive and in-depth protection.

  DeepResearch RCA: Quickly locate the root cause of issues through in-depth analysis of logs, traces, and code.

  Visualized Evidence Chain: Fully visualize the diagnostic process and evidence chain, making the diagnosis clear and enabling quick judgment of accuracy.

  Multi-Agent Collaboration: Collaboration among SRE-Agent, Code-Agent, ReportAgent, Vis-Agent, and Data-Agent.

  Open and Open-Source Architecture: OpenDeRisk is built with a completely open and open-source architecture, allowing related frameworks and code to be used out of the box in open-source projects.

  OpenDeRisk has been open-sourced on Github. The report is TBD.

  
  

•   Code Graph Model (CGM)

  CGM is a graph-based framework for software engineering tasks that employs a repository-level code graph to capture structural context and a four-step R4 chain (Rewriter, Retriever, Reranker, Reader) to resolve issues. Starting with a rewritten query from the original issue, CGM retrieves a code subgraph, prioritizes critical files via reranking, and generates a code patch using a graph-aware Reader, combining heuristic retrieval with machine learning for efficient, context-aware solutions.

  We are pleased to announce the updated version of the CGM-72B-V1.2. This new model has achieved a remarkable 44.00% resolve rate on the SWE-Bench-Lite leaderboard, marking a significant improvement in its performance and capabilities.

  CGM has been open-sourced on Github and HuggingFace. The report is on https://arxiv.org/abs/2505.16901 (accepted by NeurIPS'25).

  
  

• Ling-Coder-Lite
  
  We have open-sourced our latest MoE code LLM, Ling-Coder-Lite, available on GitHub and HuggingFace. This model is based on the Ant Ling-MoE architecture. Additionally, the technical report https://arxiv.org/abs/2503.17793 , SFT dataset and DPO dataset are also openly accessible.
  

• 
  CodeFuse is an advanced pretrained code Large Language Model (LLM) developed by our team at AntGroup. Designed to integrate seamlessly into real-world industrial scenarios, CodeFuse aims to significantly enhance the effectiveness and efficiency of software development. By leveraging the power of LLMs, CodeFuse provides developers with intelligent code suggestions, automates repetitive tasks, and offers deep insights into code quality and best practices. To foster innovation and collaboration, we have open-sourced CodeFuse, making it accessible to the broader developer community.
  
  
  CodeFuse has been open-sourced on GitHub and HuggingFace. The technical report has been published on https://arxiv.org/abs/2310.06266.
  

• MindSpore/AKG
  MindSpore/AKG also known as the Automatic Kernel Generator, is a component of the MindSpore framework, which is an open-source machine learning (ML) platform developed by Huawei. AKG plays a crucial role in optimizing the performance of neural network operations on various hardware platforms, including CPUs, GPUs, and specialized AI accelerators like Ascend. GitHub  and Gitee. The reports have been published on MICRO'20, PLDI'21 and TOCS'23
  

• 
  SVF (Static Value-Flow Analysis Framework) is a prominent open-source tool designed for static program analysis. It is widely used in software engineering and programming language research to analyze how data values flow throughout a program. By providing powerful static analysis capabilities, SVF enables developers and researchers to better understand program behavior, detect software vulnerabilities, optimize compilers, and improve program verification. GitHub.