PeopleofSuperintelligence

Builder

Yifan Shi

Also known as Shigma

Yifan Shi is a software developer whose work connects plugin design with AI systems. In 2023, he wrote a Koishi design article called “Reversible Plugin System”; from 2024 to 2025, he coauthored SWE-bench-java and reports on DeepSeek-V3, DeepSeek-R1 and DeepSeek-V3.2. His 2026 Cordis paper formalizes dynamic composition for plugin systems and supports DeepSeek Harness, and he coauthored DeepSeek-V4.1-Flash research on KV-cache compression. As of October 2026, he is a member of DeepSeek.

Now at
DeepSeek

Timeline

  1. 2023

    Reversible Plugin System

    In 2023, he writes “Reversible Plugin System” for the official Koishi documentation, laying out plugin ideas that later appear in his Cordis paper.

  2. August 2024

    SWE-bench-java and Java issues

    In August 2024, he coauthors SWE-bench-java, a Java benchmark for evaluating how language models resolve GitHub issues. The team releases its dataset with a Docker-based evaluation environment and leaderboard.

  3. December 2024

    DeepSeek-V3 technical report

    In December 2024, he coauthors the DeepSeek-V3 technical report, which describes auxiliary-loss-free load balancing and multi-token prediction as training techniques for a large language model.

  4. January 2025

    DeepSeek-R1 reasoning research

    In January 2025, he coauthors the DeepSeek-R1 report, which shows how pure reinforcement learning can prompt language models to develop reasoning patterns such as verification and self-reflection without human-labeled reasoning traces.

  5. December 2025

    DeepSeek-V3.2 sparse attention

    In December 2025, he coauthors DeepSeek-V3.2, whose DeepSeek Sparse Attention mechanism reduces computational complexity in long-context settings while preserving model performance.

  6. August 2026

    Formalizing Cordis

    In August 2026, he is first author on “A Programming Paradigm for Spatiotemporal Composability” with Wei Zhang and Tianyi Cui, under Peking University and DeepSeek-AI affiliations. The paper formalizes Cordis as a dynamic-composition framework, which supplies the plugin system for DeepSeek Harness.

  7. September 2026

    KV cache compression

    In September 2026, he coauthors DeepSeek-V4.1-Flash, a multimodal model for long-context agent workloads. It combines cache reuse and FP4 caching to reduce global KV-cache footprint and supports contexts of up to one million tokens.

  8. As of October 2026

    Member of DeepSeek

    As of October 2026, he is a member of DeepSeek.

Career

  1. Since 2026DeepSeekMember
  2. 2026Peking University

From the biography.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes DeepSeek. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes DeepSeek. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  3. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staffIncludes DeepSeek. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staff. We test the estimate against 3,913 named model contributors and dated all-staff records across Google DeepMind, OpenAI and Anthropic, plus six additional frontier-grade labs.
  4. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes DeepSeek. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.

Colleagues

People with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Yifan Shi yet. Send us a free-licensed photo.

  1. DeepSeek's “self-evolution” blueprint has been unveiled eu.36kr.com. Checked 21 October 2026.
  2. SWE-bench-java: A GitHub Issue Resolving Benchmark for Java arxiv.org. Checked 21 October 2026.
  3. DeepSeek-V3 Technical Report arxiv.org. Checked 21 October 2026.
  4. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning arxiv.org. Checked 21 October 2026.
  5. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models arxiv.org. Checked 21 October 2026.
  6. A Programming Paradigm for Spatiotemporal Composability arxiv.org. Checked 21 October 2026.
  7. DeepSeek's innovative harness treats everything as a plug-in assets.theregister.com. Checked 21 October 2026.
  8. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 21 October 2026.