About

Reliable AI, from mechanism to embodiment

I am a Ph.D. student at the Institute of Automation, Chinese Academy of Sciences, in the State Key Laboratory of Multimodal Artificial Intelligence Systems, advised by Prof. Xu-Yao Zhang and Prof. Cheng-Lin Liu. I am currently also a research intern at Kuaishou.

My research asks why large multimodal models and agents fail, and how to make them reliable. I open up their internal computation (information flow, attention, and positional structure) to find where hallucination, repetition, and spatial confusion begin. I then turn these mechanistic insights into principled fixes at both training and inference time. I carry the same reliability goals into self-evolving agents and embodied systems, from 3D spatial perception to vision–language–action (VLA) and world-action models.

Most of my work is open source; code links are listed with each paper on the Publications page.

  • Mechanistic Interpretability
  • Multimodal LLMs
  • Diffusion MLLMs
  • Self-Evolving Agents
  • Embodied AI

01News

All news →

  1. Joined Kuaishou as a Research Intern.

  2. CoTA++, the extended version of our ICLR 2026 paper, is on arXiv. It traces the repeat curse in diffusion MLLMs to three information-flow disruptions and fixes each one without training.

  3. Context-Aware Cluster Decoding, on semantic anchors for coherent diffusion MLLMs, is accepted to the EMNLP 2026 main conference.

  4. New preprint: when do the skills of self-evolving agents actually help? Coalition-Aware Skill Reliability is on arXiv.

  5. Embodied chain-of-thought via interleaved reasoning is accepted to IROS 2026 as an Oral.

02Research at a Glance

Overview →

Diagnose

Interpretability

Trace information flow inside multimodal models to find where hallucination, repetition, and positional bias begin.

Repair

Reliability

Fix failures by design: positional encodings and attention structure in training, decoding and token-level intervention at inference.

Deploy

Agency

Carry reliability into self-evolving agents and embodied systems that perceive, reason, and act.

03Selected Papers

All papers →

  • Context Tokens are Anchors: Understanding the Repeat Curse in dMLLMs from an Information Flow Perspective

    Qiyan Zhao*, Xiaofeng Zhang*, et al.

    ICLR 2026PaperCode

  • Hallucination Begins Where Saliency Drops

    Xiaofeng Zhang, …, Qiyan Zhao, et al.

    ICLR 2026OralPaperCode

  • SoPE: Spherical Coordinate-Based Positional Embedding for Enhancing Spatial Perception of 3D LVLMs

    Koon-Ting Yip*, Qiyan Zhao*, et al.

    CVPR 2026Paper

  • MCA-LLaVA: Manhattan Causal Attention for Reducing Hallucination in Large Vision-Language Models

    Qiyan Zhao*, Xiaofeng Zhang*, et al.

    ACM MM 2025PaperCode

* equal contribution

04Service

Reviewer for TPAMI, NeurIPS, ICML, ICLR, CVPR, ECCV, ACM MM, AAAI, and ACL Rolling Review.