PeopleofSuperintelligence

Researcher

Zekun Wang

Zekun Wang studies language-model attention and efficient vision-language models. He coauthored Molweni in 2020, a machine reading comprehension dataset with discourse structure over multiparty dialog, then was first author of DiDE in 2022 and SmartTrim in 2024, with an affiliation at the Harbin Institute of Technology Research Center for Social Computing and Information Retrieval. In May 2025 he was a core contributor to the Qwen3 technical report, and Gated Attention for Large Language Models, coauthored from the Qwen team at Alibaba Group, received a NeurIPS 2025 Best Paper award. As of October 2026 he is with the Qwen team at Alibaba Inc., and SlimQwen compresses Qwen3-Next-80A3B to a 23A2B model.

Works
3
Citations
0
h-index
0

Publication numbers from OpenAlex, as of 17 October 2026.

Timeline

  1. 2020

    Molweni at COLING

    In 2020 he coauthors Molweni, a machine reading comprehension dataset with discourse structure built over multiparty dialog, at the 28th International Conference on Computational Linguistics. The corpus includes 10,000 dialogs and 30,066 questions. His affiliation is the Harbin Institute of Technology Research Center for Social Computing and Information Retrieval.

  2. 2022

    Distilling a dual encoder

    DiDE, Distilled Dual-Encoder Model for Vision-Language Understanding, appears at the 2022 Conference on Empirical Methods in Natural Language Processing, with him as first author. The framework distills a fusion-encoder teacher into a dual-encoder student, with a 1% drop and 4 times faster inference.

  3. 2024

    Pruning tokens and heads

    SmartTrim, on adaptive token and attention pruning, appears at the 2024 Joint International Conference on Computational Linguistics, with him as first author. The method prunes redundant token representations and attention heads and accelerates the original model by 2-3 times, with minimal performance degradation.

  4. May 2025

    A sigmoid gate on attention

    In May 2025 he coauthors Gated Attention for Large Language Models from the Qwen team at Alibaba Group. A comparison of 30 variants of 15B mixture-of-experts models and 1.7B dense models, trained on a 3.5 trillion token dataset, finds that a head-specific sigmoid gate after scaled dot-product attention improves performance.

  5. May 2025

    Core contributor on Qwen3

    In May 2025 he is a core contributor to the Qwen3 technical report. The series includes dense and mixture-of-expert models with parameter scales from 0.6 to 235 billion, and text in 119 languages and dialects.

  6. 2025

    A NeurIPS best paper

    Gated Attention for Large Language Models receives a NeurIPS 2025 Best Paper award. The most effective SDPA output gating is used in the Qwen3-Next models.

  7. As of October 2026

    Qwen team at Alibaba

    He is with the Qwen team at Alibaba Inc. In May 2026, SlimQwen compresses Qwen3-Next-80A3B to a 23A2B model that retains competitive performance.

Sources and credits

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

  1. Molweni: A Challenge Multiparty Dialogues-based Machine Reading Comprehension Dataset with Discourse Structure - ACL Anthology aclanthology.org. Checked 16 October 2026.
  2. Zekun Wang csauthors.net. Checked 16 October 2026.
  3. Distilled Dual-Encoder Model for Vision-Language Understanding - ACL Anthology aclanthology.org. Checked 16 October 2026.
  4. SmartTrim: Adaptive Tokens and Attention Pruning for Efficient Vision-Language Models - ACL Anthology aclanthology.org. Checked 16 October 2026.
  5. Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free arxiv.org. Checked 16 October 2026.
  6. Qwen3 Technical Report arxiv.org. Checked 16 October 2026.
  7. NeurIPS 2025 Awards neurips.cc. Checked 16 October 2026.
  8. NeurIPS Poster Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free neurips.cc. Checked 16 October 2026.
  9. SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training arxiv.org. Checked 16 October 2026.