Hi! I'm a first-year master's student at Renmin University of China (RUC) and a member of AIM3 Lab. Previously, I studied Data Science at Huazhong University of Science and Technology (HUST).
I'm interested in AI agents and multimodal large language models, and enjoy building things that solve real-world problems.
I am interested in how language models reason, interact with their environments, and solve practical tasks.
NeurIPS, 2026
project page arXiv PDF dataset model (4B) model (8B)
We introduce EventLens, a reinforcement learning framework for recovering event boundaries, progression, and relations in video, and use multi-teacher on-policy distillation to combine specialized policies into one model.
arXiv, 2025
We introduce LIBERO-PRO, a benchmark that tests VLA models under changes to objects, initial states, instructions, and environments to evaluate generalization beyond memorization.
NeurIPS, 2025
We investigate backdoor vulnerabilities in VLA models through objective-decoupled optimization, revealing how triggered control deviations can coexist with clean-task performance.
arXiv, 2025
We introduce LaRMA to study when reasoning helps language model agents, comparing LLMs and large reasoning models across tool use, planning, and problem solving.
arXiv, 2024
We introduce the Open Causal Discovery Benchmark and an evaluation framework for comparing causal discovery methods on real-world data, including metrics for causal structure and causal effects.