Xijie Gong
龚熙杰About Me
I am a third-year undergraduate in Software Engineering at the University of Electronic Science and Technology of China (UESTC), expected to graduate in 2028.
My research focuses on foundation models and the science of large language models. I am especially interested in why model capabilities emerge and how learned representations and internal computations give rise to them. My current work examines these questions in knowledge representation, reasoning, and agentic decision-making. My long-term goal is to uncover principles that can guide the design and training of foundation models.
Research interests: Foundation Models · Science of Large Language Models · Interpretability
Recent News
Publication and Competitions
How Do Agentic LLMs Decide to Call Tools? A Scaffold Default Controlled by Suppression
Tool calling is central to agentic LLMs, yet its internal decision mechanism remains poorly understood. Using minimally contrastive prompts, we identify a localized tool-call vector with causal sufficiency and necessity. We find that the agent scaffold induces a strong prior to call tools, while analysis-oriented requests suppress this prior.
Verbalizing Multi-Token Concepts in LLMs
Concept Lens extends token-level lens methods to multi-token concepts by using token clues to propose candidate concepts and scoring their representations against the source activation. Across 2,400 multi-hop cloze prompts and five LLMs (8B–70B), it enables multi-token readout and supports causal concept-swap interventions.
The RoboSense Challenge: Sense Anything, Navigate Anywhere, Adapt Across Platforms
The RoboSense 2025 Challenge unifies five complementary tracks spanning language-grounded decision making, socially compliant navigation, sensor configuration generalization, cross-view and cross-modal correspondence, and cross-platform 3D perception. This report provides standardized datasets, baselines, and unified evaluation protocols for robust robot sensing under domain shifts, and consolidates insights from 23 winning solutions.