Publications

Publications in reverse chronological order.

2027

  1. Focus on What You Need: Global-to-Local Dynamic Attention for Multimodal Sarcasm Detection
    Dehao Wu and Xingjie Zhuang
    ICASSP 2027 Under review
    arXiv

    A global-to-local dynamic attention framework for multimodal sarcasm detection with fine-grained text-image alignment.

2026

  1. Large-scale agentic analysis identifies recurrent regulators linking diseases across patient subgroups
    Dehao Wu, Yuting Liao, Qiqi Xie, Yang Li, and Haohan Wang
    npj Systems Biology and Applications Under review

    An LLM-agent framework that builds a network of gene regulatory networks at patient-subgroup resolution, comparing 212,469 network pairs across 132 diseases to recover shared regulatory programmes.

  2. Agentic AI uncovers conserved cross-tissue protein co-abundance programs inaccessible to single-dataset analysis
    Runyu Guan*, Dehao Wu*, Qiqi Xie, Yang Li, and Haohan Wang
    Bioinformatics Under review · *Equal contribution
    arXiv

    An LLM-agent pipeline for generating and comparing protein co-abundance networks across different human tissues.

  3. Agents’ Last Exam
    Yiyou Sun, Xinyang Han, Weichen Zhang, …, Dehao Wu, …, and Dawn Song
    NeurIPS 2026
    arXiv

    A living benchmark evaluating AI agents on long-horizon, economically valuable real-world tasks with verifiable outcomes, developed with 250+ industry experts.

  4. Composing Task-specific Agent Harnesses at Test Time with Reusable Primitives
    Peng Kuang, Haibo Jin, Dehao Wu, Feiyang Deng, Xiaopeng Yuan, Jerry Wang, and Haohan Wang
    ICLR 2027 Under review
    arXiv

    STITCH mines reusable harness primitives from failed agent trajectories and composes them into a task-specific harness at test time, raising success rates by up to 12 points over fixed harnesses.

  5. Large-scale Repository Engineering via Agent-Native Reusable Code Primitives
    Haibo Jin, Peng Kuang, Xucheng Yu, Jerry Wang, Dehao Wu, and Haohan Wang
    ICLR 2027 Under review
    arXiv

    LEGO adapts reusable, tested code primitives from a 1,424-entry library to build complete repositories, improving all 13 evaluated LLM backbones on the 522-task LEGO-REPO benchmark.