PeopleofSuperintelligence

Researcher

Xin Cheng

Xin Cheng is an AI researcher whose work connects machine translation, retrieval-augmented generation, and large language models. His papers span contrastive translation memories, Selfmem's self-memory framework, and SCALE's unified translation engine. He contributes to the DeepSeek-V3 and DeepSeek-R1 papers, then coauthors work on Deep Sparse Embedding and DSpark. As of October 2026, he is affiliated with Peking University and DeepSeek-AI.

Timeline

  1. 2022

    Contrastive translation memories

    In 2022, he coauthors an EMNLP paper on neural machine translation with contrastive translation memories. Its retrieval method avoids redundancy and uninformativeness among similar translation pieces.

  2. 2023

    Self-memory for text generation

    In 2023, he coauthors the NeurIPS paper Selfmem. Its retrieval-augmented generator grows an unbounded memory pool, and a selector chooses one output as memory for the next generation round.

  3. 2024

    SCALE for machine translation

    In 2024, he coauthors SCALE, which connects a compact specialized translation model and a general-purpose large language model as one unified translation engine.

  4. December 2024

    DeepSeek-V3 research and engineering

    In December 2024, he contributes to DeepSeek-V3's Research & Engineering work. The model has 671 billion total parameters, with 37 billion active for each token.

  5. September 2025

    DeepSeek-R1 reasoning research

    In September 2025, he coauthors the DeepSeek-R1 study, which finds that pure reinforcement learning can incentivize reasoning without human-labelled reasoning trajectories. The paper describes emergent self-reflection and verification.

  6. July 2026

    Deep Sparse Embedding

    In July 2026, he coauthors a paper introducing conditional memory through Deep Sparse Embedding (DSE), which indexes a large embedding table with local n-grams. At 27 billion parameters, DSE outperforms an iso-parameter and iso-FLOPs mixture-of-experts baseline on knowledge and reasoning benchmarks.

  7. July 2026

    DSpark in DeepSeek-V4 serving

    In July 2026, he coauthors DSpark, a confidence-scheduled speculative decoding system. DeepSeek deploys it in the DeepSeek-V4 serving system under live user traffic.

  8. As of October 2026

    Affiliations at Peking and DeepSeek

    He is affiliated with Peking University and DeepSeek-AI. In September 2026, he coauthors DeepSeek-AI's technical report DeepSeek-V4.1-Flash.

Career

  1. NowPeking University
  2. NowDeepSeek

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.

Sources and credits

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

  1. Neural Machine Translation with Contrastive Translation Memories - ACL Anthology aclanthology.org. Checked 22 October 2026.
  2. Lift Yourself Up: Retrieval-augmented Text Generation with Self-Memory proceedings.neurips.cc. Checked 22 October 2026.
  3. SCALE: Synergized Collaboration of Asymmetric Language Translation Engines - ACL Anthology aclanthology.org. Checked 22 October 2026.
  4. DeepSeek-V3 Technical Report arxiv.org. Checked 22 October 2026.
  5. DeepSeek-V3 Technical Report arxiv.org. Checked 22 October 2026.
  6. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning nature.com. Checked 22 October 2026.
  7. Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models - ACL Anthology aclanthology.org. Checked 22 October 2026.
  8. DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation arxiv.org. Checked 22 October 2026.
  9. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 22 October 2026.