Zichen Tang

Second-year PhD, Computer Science and Engineering
Hong Kong University of Science and Technology

I am a second-year PhD student in Computer Science and Engineering at the Hong Kong University of Science and Technology, advised by Prof. Bo Li.

Before HKUST, I completed an MPhil in Data Science and Analytics at HKUST (Guangzhou) under Prof. Xiaowen Chu and Prof. Sihong Xie, working on communication-efficient distributed and federated learning. I received my BEng in Computer Science from HKUST with a double major in Mathematics.

email · google scholar · github · cv (pdf)

Research

Federated / Distributed Training · Data-centric LLM Training · LLM Safety

News

  • 2026-10 Paper Anchoring the Bridge Across Hops: Unlocking Latent Multi-Hop Reasoning under Knowledge Injection accepted at CL4FMAgents @ NeurIPS 2026 Workshop.
  • 2026-09 Paper BiasTrojan: LLM Judgers Are Easily Distorted by Few Hundreds of Contrastive Biased Training Data accepted at NeurIPS 2026 as a poster.
  • 2026-09 Paper Is Your LLM-as-a-Recommender Agent Trustable? LLMs' Recommendation is Easily Hacked by Biases (Preferences) accepted at the NeurIPS 2026 Evaluations and Datasets Track as a poster.
  • 2026-05 Recognized as an ICML 2026 Gold Reviewer.
  • 2026-05 Paper Capturing and Mitigating Gradient Aggregation Errors for Fault-Tolerant Distributed Training accepted at ICML 2026.
  • 2026-03 Preprint released: Is Your LLM-as-a-Recommender Agent Trustable? LLMs' Recommendation is Easily Hacked by Biases (Preferences).
  • 2026-02 Started TAing COMP 3511 (Operating Systems) at HKUST.
  • 2026-01 Paper Ghost in the Cloud: Your Geo-Distributed LLM Training is Easily Manipulated accepted at ICLR 2026.
  • 2026-01 Paper DreamDDP: Accelerating Data Parallel Distributed LLM Training with Layer-wise Scheduled Partial Synchronization accepted at MLSys 2026.
  • 2025-09 Began my PhD in Computer Science and Engineering at HKUST.

Selected Publications

  1. Overview of latent multi-hop composition, knowledge-injection rewrites, BRACE, and explicit elicitation
    CL4FMAgents @ NeurIPS'26 Workshop · 2026 · Accepted
    Anchoring the Bridge Across Hops: Unlocking Latent Multi-Hop Reasoning under Knowledge Injection
    Zichen Tang, Zhenheng Tang, Yifan Hou, Hanwen Xing, Xinda Qi, Xiaowen Chu, Bo Li
    openreview
  2. Overview of BiasTrojan: context-aware and contrastive bias injection with counterfeited reasoning, evaluation, and deployment attacks
    NeurIPS'26 (Poster) · 2026 · Accepted
    BiasTrojan: LLM Judgers Are Easily Distorted by Few Hundreds of Contrastive Biased Training Data
    Zichen Tang*, Zhenheng Tang*†, Qian Wang, Gaoning Pan, Yuhan Yang, Wei He, Shaohuai Shi, Xiaowen Chu, Bo Li†
    * equal contribution; † corresponding authors
  3. Overview of the data-synthesis pipeline with quality calibration and bias evaluation
    NeurIPS'26 Evaluations and Datasets Track (Poster) · 2026 · Accepted
    Is Your LLM-as-a-Recommender Agent Trustable? LLMs' Recommendation is Easily Hacked by Biases (Preferences)
    Zichen Tang*, Zirui Zhang*, Qian Wang, Zhenheng Tang†, Xiaowen Chu, Bo Li†
    * equal contribution; † corresponding authors
    arxiv
  4. KnownContext overview: knowledge rewriting, atomic data rewriting, a rewriting taxonomy, and the evaluation protocol
    ICML'26 Workshop · 2026
    Enhancing Knowledge Injection with Surrounding Backgrounds in Continual Training LLMs
    Zichen Tang, Z. Tang, Y. Hou, P. Dong, X. Liu, S. Shi, X. Chu, B. Li
  5. Ghost in the Cloud overview: trigger-based jailbreak injection, two defenses, and the deployment attack
    ICLR'26 · 2026 · Accepted
    Ghost in the Cloud: Your Geo-Distributed Large Language Models Training is Easily Manipulated
    Zichen Tang*, Z. Tang*, G. Pan, B. Liu, X. He, K. Lai, X. Chu, B. Li
    * equal contribution
  6. DreamDDP system overview: profiler-driven schedule pipeline with layer-wise partial synchronization
    MLSys'26 · 2026 · Accepted
    DreamDDP: Accelerating Data Parallel Distributed LLM Training with Layer-wise Scheduled Partial Synchronization
    Z. Tang*, Zichen Tang*, J. Huang, X. Pan, R. Yan, Y. Wang, A. C. Zhou, S. Shi, X. Chu, B. Li
    * equal contribution
    arxiv
  7. Per-round timeline: adaptive bandwidth-aware compression saves communication time over uniform compression
    ICPP'24 · 2024 · Published
    Bandwidth-Aware and Overlap-Weighted Compression for Communication-Efficient Federated Learning
    Zichen Tang, J. Huang, R. Yan, Y. Wang, Z. Tang, S. Shi, A. C. Zhou, X. Chu
    arxiv

→ full publication list