Kaijie Zhu
kaijiezhu@ucsb.edu
CA, US
I’m a third-year Ph.D student at UCSB, fortunately advised by Prof. William Wang and Prof. Wenbo Guo.
My research interests span evaluating, training, and securing AI agent systems.
I am expected to graduate in 2027 and am open to industry positions (e.g., research scientist / research engineer on LLM post-training and agentic RL). Feel free to reach out!
- Reasoning and Agent: training (LLM Value Functions for Agentic RL, w/ Apple, TermiGen arXiv, rePIRL ICML 2026), benchmarking (DevOps-Gym, ICLR 2026), defending indirect prompt injection (MELON, ICML 2025)
- Dynamic Evaluation of LLMs: DyVal series (ICLR’24 Spotlight, ICML’24), PromptBench (JMLR MLOSS)
- Adversarial Robustness: LLM prompt robustness (PromptRobust CCS’24 LAMPS), CNN robustness (RiFT ICCV’23)
news
| May 1, 2026 | rePIRL is accepted by ICML 2026. |
|---|---|
| Jan 26, 2026 | DevOps-Gym is accepted by ICLR 2026. |
| May 1, 2025 | MELON is accepted by ICML 2025. |
| Feb 25, 2025 | Hosting the AAAI 2025 Tutorial on Evaluating Large Language Models: Challenges and Methods with Prof. Jindong Wang, Dr. Linyi Yang, Prof. Yue Feng, and Prof. Yue Zhang. |
| Jan 20, 2025 | Selected to present a talk at the KAUST Rising Stars in AI Symposium 2025. |
selected publications
-
-
TermiGen: High-Fidelity Environment and Robust Trajectory Synthesis for Terminal AgentsarXiv, 2026 -
rePIRL: Learn PRM with Inverse RL for LLM ReasoningICML, 2026 -
-
MELON: Indirect Prompt Injection Defense via Masked Re-execution and Tool ComparisonICML, 2025 -
DyVal: Graph-informed Dynamic Evaluation of Large Language ModelsICLR (Spotlight), 2024 -
PromptBench: Towards Evaluating the Robustness of Large Language Models on Adversarial PromptsCCS LAMPS Workshop, 2023 -
Improving Generalization of Adversarial Training via Robust Critical Fine-TuningICCV, 2023 -
DyVal 2: Dynamic Evaluation of Large Language Models by Meta Probing AgentsICML, 2024
