Ph.D. Student · Shanghai Jiao Tong University
Email: xyz9911 [at] sjtu.edu.cn
I study how embodied agents build internal models of their world - remembering what they have seen, imagining what comes next, and coordinating with other agents to act on it. I draw on how the brain solves the same problems, especially the hippocampal link between memory, simulation, and spatial cognition, and use these ideas to build navigation agents that are more capable, efficient, and interpretable. I also work on language models, multimodal LLMs and LLM agents.
I'm a final-year Ph.D. student in the School of Computer Science at SJTU. I lead projects end to end, from choosing the research direction and idea through implementation, experiments, and paper writing, all on a single A100 GPU.
On the job market. I'm looking for postdoctoral positions starting 2027, especially at the intersection of AI and computational neuroscience.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2026
A world model imagines future states and uses them to query long-term memory of past observations and behaviours, so the agent improves over time without extra supervision. +5.4 SPL over the strongest baseline, with 8.3× faster training and 74% less inference memory.
AAAI Conference on Artificial Intelligence, 2025 Oral
The first to adapt a multimodal LLM into an outdoor navigation agent, via a three-phase tuning recipe and an LLM-driven synthetic data pipeline. +7.3 points over the previous state of the art, with the entire training done in 14 hours on a single A100 (vs. 672 GPU-hours reported for indoor-VLN work NaVid).
Preprint
The first systematic formalization of multi-agent VLN as constrained coordination; a verified four-stage LLM pipeline yields 11,724 episodes across 145 scenes with teams of up to 4 agents. Our system couples a compact LLM scheduler, a shared topological memory, and conflict-aware execution, and its planner needs only 2.1 GB GPU memory at inference and 21.1 GB for training.
IEEE Transactions on Audio, Speech, and Language Processing (TASLP), 2026
Finds knowledge gaps and injects domain knowledge into LLMs, with adaptive pruning. ~6% average gain across 15 financial, legal, and medical benchmarks with up to 29% fewer inference tokens.