Xijie Gong
龚熙杰About Me
I am a third-year undergraduate student in Software Engineering at the University of Electronic Science and Technology of China (UESTC), expecting to graduate in 2028. I am currently a research intern at the Delta-I Lab, College of AI (CAI), Tsinghua University. Passionate about large language models and agents, I plan to pursue a Ph.D. and continue advancing research in this field.
My research centers on a fundamental question: how do foundation models work? I study how capabilities such as knowledge, reasoning, and agentic behavior arise from the representations and computations within large language models. I am particularly interested in understanding why these capabilities emerge and how they are realized by the model's internal computation. Ultimately, I hope that a deeper scientific understanding of large language models can reveal principles for building better foundation models.
Research interests: Foundation Models · Science of Large Language Models · Interpretability
Recent News
Publications & Preprints
How Do Agentic LLMs Decide to Call Tools? A Scaffold Default Controlled by Suppression
Tool calling is central to agentic LLMs, yet the decision mechanism remains poorly understood. We propose minimal contrastive pairs of agentic prompts, isolate a single causally necessary and sufficient decision vector, and demonstrate that tool-use is a scaffold-induced default that non-tool verbs suppress.
Verbalizing Multi-Token Concepts in LLMs
Lens methods map activations to single tokens, missing multi-token human concepts. We introduce Concept Lens: token clues guide candidate search, the model scores representations against activations. Evaluated across 2,400 multi-hop clozes on five LLMs (8B–70B) with causal concept swaps.
The RoboSense Challenge: Sense Anything, Navigate Anywhere, Adapt Across Platforms
A multi-platform benchmark evaluating general-purpose 3D spatial perception, LiDAR-inertial-visual sensor fusion, and zero-shot platform adaptation.