Nan Jiang

Research

Two main threads: efficient diffusion language models, and neuro-symbolic AI for scientific discovery.

Efficient Inference for Diffusion Language Models

Goal: Reduce the inference cost of diffusion-based LLMs for language generation.

Strategies:

  • Reduce the number of diffusion iterations via distillation.
  • Develop a sparse attention mechanism for long-sequence generation.

Related publications

  1. Preprint
    PreSTO: Predictive Subtree Prefetching for Fast LLM Power Sampling
    Nan Jiang, Panagiotis Theodoropoulos, Wei Deng
    Preprint
    arXivwebsitecode
  2. Preprint
    Explore Broadly, Reason Sharply: Push Small Models toward the Frontier via Sampling
    Panagiotis Theodoropoulos, Nan Jiang, Xintong Duan, Ali Hasan, Yuriy Nevmyvaka, Evangelos A. Theodorou, Wei Deng
    Preprint
    arXiv
  3. ICLR
    Ultra-Fast Language Generation via Discrete Diffusion Divergence Instruct
    Haoyang Zheng, Xinyang Liu, Cindy Xiangrui Kong, Nan Jiang, Zheyuan Hu, Weijian Luo, Wei Deng, Guang Lin
    ICLR 2026
    OpenReviewwebsitecode
  4. AISTATS
    DIVERSED: Relaxed Speculative Decoding via Dynamic Ensemble Verification
    Ziyi Wang, Siva Rajesh Kasa, Ankith M S, Santhosh Kumar Kasa, Jiaru Zou, Sumit Negi, Ruqi Zhang, Nan Jiang, Qifan Song
    AISTATS 2026
    OpenReviewpostercode
  5. TMLR

Neuro-Symbolic AI for Scientific Discovery

Goal: Accelerate the discovery of physical knowledge from experimental data.

Strategies:

  • Design control-variable experiments.
  • Perform active data querying guided by phase portraits.
  • Learn a symbolic-equivalence–aware objective.

Related publications

  1. ICLR
  2. AAAI
  3. ICML
  4. IJCAI
  5. AAAI
  6. ECML-PKDD
  7. Springer
    Expediting Symbolic Regression for Science Using Scientific Approaches
    Md Nasim, Nan Jiang, Yexiang Xue
    Book Chapter, Computational Approaches to Scientific Discovery (Springer)

Related: see the Genesis Mission for context on AI-driven scientific discovery. All papers are on the publications page.