Agentic AI
Agents that plan, keep memory across long interactions, act in the world, and improve themselves from feedback.
Chengzhi Liu
My research focuses on Multimodal Agentic, Trustworthy AI, and Recursive Self-Improvement.
I am a Ph.D. student in Computer Science at the University of California, Santa Barbara, advised by Prof. Xin Eric Wang. I am also a Student Researcher at Google Cloud AI, where I have worked since June 2026. Previously, I studied Information and Computing Science at the University of Liverpool.
I study how multimodal agents perceive, reason, and act in dynamic environments, with an emphasis on memory, long-horizon interactions, and learning from feedback. I am particularly interested in recursive self-improvement: enabling agents to refine their reasoning and behavior through experience, while remaining reliable, safe, and grounded in what they observe.
My work has appeared at NeurIPS, ICLR, CVPR, AAAI, COLM, and ACL.
Ph.D., since 2025
Cloud AI, Jun 2026 – present
B.Sc., 2025
Research
Multimodal Perception & Reasoning
Visual grounding, reasoning in latent space, streaming video understanding, and keeping models faithful to what they see.
Trustworthy AI
Safety evaluation for multimodal models and agent harnesses, hallucination mitigation, and alignment of reasoning models.
News
Three papers accepted to NeurIPS 2026: WorldMemArena, Survive or Collapse, and Agents’ Last Exam.
Started as a Student Researcher at Google Cloud AI.
Latent-space multimodal reasoning (DMLR) accepted to CVPR 2026.
EvoPresent accepted to ICLR 2026, and selected as a Spotlight at the ICLR 2026 RSI Workshop.
Invited talk on multimodal latent-space reasoning at the NLP Academic Exchange Platform.
Started my Ph.D. in Computer Science at UC Santa Barbara.
Papers at NeurIPS, ICLR, AACL, CVPR (Oral), and AAAI (Oral) on reasoning hallucination, multimodal safety, and medical imaging.
Scroll for earlier updates
Publications
* Equal contribution. Full list on Google Scholar
First & co-first author
StreamMind: Proactive Video Agents for Evolving User Intent
Video agents that decide when to speak up as a user's intent shifts during a live stream.
WorldMemArena: Evaluating Multimodal Agent Memory Through Action–World Interaction
A benchmark that tests whether multimodal agents remember what their own actions did to the world.
Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space
Interleaving visual and textual reasoning directly in latent space, without explicit tool calls.
Auditing Agent Harness Safety
Auditing the safety risks introduced by the harnesses that run AI agents.
Presenting a Paper is an Art: Self-Improvement Aesthetic Agents
Agents that generate paper presentations and improve their own visual design through aesthetic feedback.
More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models
Longer reasoning chains can pull multimodal models away from the image and amplify hallucination.
Seeing Far and Clearly: Mitigating Hallucinations in MLLMs
Reshaping attention so multimodal models keep looking at distant visual tokens, reducing hallucination.
Multimodal Situational Safety
The same request can be safe or unsafe depending on the scene; MSSBench tests whether models notice.
The Hidden Risks of Large Reasoning Models: A Safety Assessment of R1
A safety assessment of open reasoning models, including how risky content surfaces in the thinking process.
Incomplete Modality Disentangled Representation for Ophthalmic Disease Grading and Diagnosis
Disentangled representations that stay robust when some imaging modalities are missing.
Collaborative work
Survive or Collapse: The Asymmetric Roles of Data Gating and Reward Grounding in Self-Play RL
Why some self-play RL runs keep improving while others collapse, traced to data gating and reward grounding.
Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling
Token-level value pretraining for predicting and controlling generation length in language and vision-language models.
Intervening Anchor Token: Decoding Strategy in Alleviating Hallucinations for MLLMs
A decoding strategy that intervenes on anchor tokens to reduce hallucination.
No papers match that search. Try a title word, an author, or a venue such as “CVPR”.
Experience
-
Google Cloud AI
Student Researcher
-
University of California, Santa Barbara
Ph.D. in Computer Science, advised by Xin Eric Wang
-
The University of Liverpool
B.Sc. in Information and Computing Science
Academic Services
- Conference Reviewer: NeurIPS (2026), ICML (2026), CVPR (2026), ICLR (2026, 2027), AAAI (2026, 2027), AAAI-AIA (2026).
- Workshop Reviewer: MICCAI SAFER (2026), COLM WAB (2026).
Invited Talks
- Multimodal latent-space reasoningNLP Academic Exchange Platform
- Canvas: aesthetic-driven automatic presentation generation
Open to collaborations on multimodal agentic AI, trustworthy AI, and recursive self-improvement.
chengzhi@ucsb.edu