|
Zijun Wang I am a PhD student in Computer Science and Engineering at University of California, Santa Cruz (UCSC), advised by Prof. Cihang Xie in the VLAA Lab. I received my Bachelor of Engineering in Computer Science and Technology from Zhejiang University. My research focuses on AI agents for automated AI research, with the long-term goal of recursive self-improvement (RSI). My interests also include data-centric LLM training and AI safety. I study how agents learn from experimental experience to predict outcomes, select promising experiments, and propose improvements to AI systems. My recent work explores research world models and reinforcement learning for model-improvement proposals. I am currently interning at Amazon in Palo Alto, working on RSI agents and research world models with Hanqing Lu. I am always open to research discussions and collaborations : ) |
|
News[2026.10] New work on AI research agents: Research World Models for predicting experimental outcomes, and LM2LM for learning to propose model improvements. [2026.10] ALPS is out: visual state encoders for world models. [2026.08] Your Agent, Their Asset and Harness Updating Is Not Harness Benefit are accepted by EMNLP 2026. [2026.07] Chasing the Public Score is accepted by COLM 2026. [2026.06] I joined Amazon in Palo Alto as an intern, working on RSI agents and research world models with Hanqing Lu. [2026.06] VisualClaw is out! A real-time, personalized agent for the physical world. [2026.04] NAG and InfoLaw are accepted by ICML 2026 [2026.04] We release VLAA-GUI, the verified best open-source GUI systems on OSWorld! [2026.04] CIK-Bench is out! A real-world safety analysis exposing architectural vulnerabilities of personal AI agents. [2026.02] MIRA is accepted by CVPR 2026 [2025.11] STAR-1 is accepted by AAAI 2026 (oral) [2025.06] I'm joining ByteDance as Research Scientist Intern. [2025.04] STAR-1 is out! Using just 1K data to make your reasoning LLMs much safer. [2025.01] AttnGCG is accepted by TMLR 2025 [2024.12] Graph Patches is accepted by KDD 2025 [2024.07] Unicorns is accepted by ECCV 2024 [2023.12] Second Place in both base & large model subtracks of Red Teaming LLM@NeurIPS 2023, Trojan Detection Challenge [Code] |
Publications |
2026 |
|
|
Language Models as AI Research World Models
Research world models predict experimental outcomes and guide experiment selection using prior research experience. |
|
Language Models Learn to Improve Language Models
LM2LM trains models to propose better model improvements with reinforcement learning, using research-world-model feedback calibrated by real experiments. |
|
Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
|
 |
Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw
|
 |
Target-Oriented Pretraining Data Selection via Neuron-Activated Graph
|
 |
STAR-1: Safer Alignment of Reasoning LLMs with 1K Data
|
|
VisualClaw: A Real-Time, Personalized Agent for the Physical World
|
 |
VLAA-GUI: Knowing When to Stop, Recover, and Search, A Modular Framework for GUI Automation
|
 |
When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought
|
|
ALPS: Visual State Encoders for World Models
A visual state encoder unifying physical-state information, language alignment, and pixel reconstruction for world models. |
 |
InfoLaw: Information Scaling Laws for Large Language Models with Quality-Weighted Mixture Data and Repetition
|
 |
Chasing the Public Score: User Pressure and Evaluation Exploitation in Coding Agent Workflows
|
|
ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning
An agentic framework for actively seeking multimodal clinical evidence from patient records and imaging tools. |
 |
Mimicking the Physicist's Eye: A VLM-centric Approach for Physics Formula Discovery
|
2025 |
|
 |
AttnGCG: Enhancing Jailbreaking Attacks on LLMs with Attention Manipulation
|
 |
AHELM: A Holistic Evaluation of Audio-Language Models
|
 |
Where on Earth? A Vision-Language Benchmark for Probing Model Geolocation Skills Across Scales
|
 |
Handling Feature Heterogeneity with Learnable Graph Patches
|
2024 |
|
 |
How Many Unicorns Are in This Image? A Safety Evaluation Benchmark for Vision LLMs
|
|
Robust and Efficient AI-Based Attack Recovery in Autonomous Drones
An autonomous attack-recovery architecture using common-sense reasoning to plan recovery actions for drones. |
Experience |
|
Aug. 2024 - Present, VLAA Lab, UC Santa Cruz PhD student in Computer Science and Engineering, advised by Prof. Cihang Xie. Expected graduation: Jun. 2028. Aug. 2023 - Aug. 2024, VLAA Lab, UC Santa Cruz Visiting Research Intern advised by Prof. Cihang Xie, Adversarial Attacks on LLMs & VLLMs |
|
Jun. 2026 - Present, Amazon, Palo Alto, CA Intern advised by Hanqing Lu, working on RSI agents and research world models. Co-first-authored Language Models as AI Research World Models and co-authored Language Models Learn to Improve Language Models (arXiv 2026), as well as Harness Updating Is Not Harness Benefit (EMNLP 2026). |
|
Jun. 2025 - Apr. 2026, ByteDance, San Jose, CA Research Scientist Intern advised by Fengze Liu, working on foundation LLM pretraining. First-authored NAG for target-oriented pretraining data selection and co-authored InfoLaw for data-aware scaling laws; both accepted by ICML 2026. |
|
Jan. 2023 - Jul. 2023, Zhejiang University Research Assistant advised by Prof. Yang Yang, Generalized Graph Pre-training Sep. 2020 - Jun. 2024, Zhejiang University Bachelor of Engineering in Computer Science and Technology, GPA: 3.92/4.0 |
Awards- Second Place in both base and large model subtracks of the NeurIPS 2023 Trojan Detection Challenge, Red Teaming LLM track; team leader. [Code] - National Scholarship (top 0.2% nationwide), Ministry of Education of the People's Republic of China, 2021 - First-class Scholarship of Zhejiang University - Provincial Government Scholarship of Zhejiang Province |
|
Last Update 2026.11.05. Thanks to Jon Barron. |