Zikui Cai

I am currently at Amazon, where I work on robotics with a focus on mobility for humanoid robots. Previously, I was a postdoctoral associate at University of Maryland, College Park, working with Furong Huang and Tom Goldstein. I obtained my Ph.D. in Electrical and Computer Engineering at the University of California, Riverside, under the guidance of M. Salman Asif. Originally from Wuhan, the "River City" of China, I completed my B.S. in Automation at Wuhan University of Technology.

My research focuses on understanding and enabling reliable long-horizon behavior in autonomous agents and robotic systems, with an emphasis on mobility, memory, adaptation, and robustness through interaction with the environment.

Research Interests

World Modeling and Multimodal Reasoning

Safety, Alignment, and Model Adaptation

  • Agentic safety and alignment in multi-agent ecosystems [ AegisLLM ]
  • Controllable and safe AI via model editing and unlearning [ SLUG, Robust Eval ]
  • Enhancing user privacy via de-identification while preserving utility [ Disguise ]

Robust Perception and Decision-Making

News

  • [Jul. 2026]Excited to join Amazon to work on robotics.
  • [Jul. 2026]Beyond Episodic Evaluation on sequential memory for embodied question answering has been accepted to IROS 2026 @ Pittsburgh.
  • [Mar. 2026]Low-Frequency Trap on VLM Temporal Capability Profiling has been accepted to ICLR 2026 ICBNB @ Rio de Janeiro.
  • [Feb. 2026]TraceGen on 3D Trace Generation for Robotic Manipulation has been accepted to CVPR 2026 @ Denver.
  • [Jan. 2026]Zebra-CoT on Interleaved Vision-Language Reasoning has been accepted to ICLR 2026 @ Rio de Janeiro.
  • [Jun. 2025]Our paper on Efficient Training for VLM has been accepted to ICCV 2025 @ Honolulu.
  • [Jun. 2025]Our paper on Model Tampering Attack for Robust Evaluation has been accepted to TMLR.
  • [May 2025]Our paper SLUG on Efficient Unlearning has been accepted to ICML 2025 @ Vancouver.
  • [Mar. 2025]AegisLLM on Agentic System for Safety has been accepted to ICLR 2025 BuildingTrust @ Singapore.
  • [Oct. 2024]Our papers on Efficient and Robust Unlearning are accepted to NeurIPS 2024 SafeGenAI @ Vancouver.

Selected Publications

Beyond Episodic Evaluation: Memory Architectural Bottlenecks in Sequential Embodied Question Answering
Zikui Cai*†, Kaushal Janga*, Tan Dat Dao*, Seungjae Lee*, Shivin Dass, Mingyo Seo, Kaiyu Yue, Mintong Kang, Nandhu Pillai, Monte Hoover, Aadi Palnitkar, Ruchit Rawal, Ruijie Zheng, Bo Li, Yuke Zhu, Roberto Martín-Martín, Tom Goldstein, Furong Huang
IROS 2026
arxiv project code
The Low-Frequency Trap: Why Scaling Doesn't Solve Simple Temporal Counting
Sarvesh Baskar*, Muhammad R. Islam*, Zikui Cai†, Ankit Nakhawa, Anirudh Satheesh, Tom Goldstein, Furong Huang
ICLR 2026 ICBNB
paper
TraceGen: World Modeling in 3D Trace Space Enables Learning from Cross-Embodiment Videos
Seungjae Lee, Yoonkyo Jung, Inkook Chun, Yao-Chih Lee, Zikui Cai, Hongjia Huang, Aayush Talreja, Tan Dat Dao, Yongyuan Liang, Jia-Bin Huang, Furong Huang
CVPR 2026
arxiv code
Zebra-CoT: A Dataset for Interleaved Vision-Language Reasoning
Ang Li*, Charles Wang*, Deqing Fu*, Kaiyu Yue*, Zikui Cai*, Wang Bill Zhu*, Ollie Liu*, Peng Guo*, Willie Neiswanger, Furong Huang, Tom Goldstein, Micah Goldblum
ICLR 2026
arxiv dataset
Zero-Shot Vision Encoder Grafting via LLM Surrogates
Kaiyu Yue, Vasu Singla, Menglin Jia, John Kirchenbauer, Rifaa Qadri, Zikui Cai, Abhinav Bhatele, Furong Huang, Tom Goldstein
ICCV 2025
arxiv code
Targeted Unlearning with Single Layer Unlearning Gradient
Zikui Cai, Yaoteng Tan, M Salman Asif
ICML 2025
arxiv code
AegisLLM: Scaling Agentic Systems for Self-Reflective Defense in LLM Security
Zikui Cai, Shayan Shabihi, Bang An, Zora Che, Brian R. Bartoldson, Bhavya Kailkhura, Tom Goldstein, Furong Huang
ICLR 2025 BuildingTrust
arxiv code
Model Tampering Attacks Enable More Rigorous Evaluations of LLM Capabilities
Zora Che, Stephen Casper, Robert Kirk, Anirudh Satheesh, Stewart Slocum, Lev E McKinney, Rohit Gandikota, Aidan Ewart, Domenic Rosati, Zichu Wu, Zikui Cai, Bilal Chughtai, Yarin Gal, Furong Huang, Dylan Hadfield-Menell
TMLR 2025
arxiv code
Can Textual Unlearning Solve Cross-Modality Safety Alignment?
Trishna Chakraborty, Erfan Shayegani, Zikui Cai, Nael Abu-Ghazaleh, M. Salman Asif, Yue Dong, Amit K. Roy-Chowdhury, Chengyu Song
EMNLP 2024
arxiv
Disguise without Disruption: Utility-Preserving Face De-Identification
Zikui Cai, Zhongpai Gao, Benjamin Planche, Meng Zheng, Terrence Chen, M. Salman Asif, Ziyan Wu
AAAI 2024
arxiv
Ensemble-based Blackbox Attacks on Dense Prediction
Zikui Cai*, Yaoteng Tan*, M. Salman Asif
CVPR 2023
arxiv code
Blackbox Attacks via Surrogate Ensemble Search
Zikui Cai, Chengyu Song, Srikanth V. Krishnamurthy, Amit K. Roy-Chowdhury, M. Salman Asif
NeurIPS 2022
arxiv code poster slides talk
Zero-Query Transfer Attacks on Context-Aware Object Detectors
Zikui Cai, Shantanu Rane, Alejandro E. Brito, Chengyu Song, Srikanth V. Krishnamurthy, Amit K. Roy-Chowdhury, M. Salman Asif
CVPR 2022
arxiv open access
Context-Aware Transfer Attacks for Object Detection
Zikui Cai, Xinxin Xie, Shasha Li, Mingjun Yin, Chengyu Song, Srikanth V. Krishnamurthy, Amit K. Roy-Chowdhury, M. Salman Asif
AAAI 2022
arxiv code slides poster talk
Exploiting Multi-Object Relationships for Detecting Adversarial Attacks in Complex Scenes
Mingjun Yin, Shasha Li, Zikui Cai, Chengyu Song, M. Salman Asif, Amit K. Roy-Chowdhury, Srikanth V. Krishnamurthy
ICCV 2021
arxiv open access
Data-Driven Illumination Patterns For Coded Diffraction Imaging
Zikui Cai, Rakib Hyder, M. Salman Asif
ICIP 2021
paper
Solving Phase Retrieval with a Learned Reference
Rakib Hyder*, Zikui Cai*, M. Salman Asif
ECCV 2020
arxiv code