Jiawei Gu （顾嘉炜）

An evaluation framework where an off-the-shelf VLM drives a full 3D human body by issuing atomic skill commands, so its decisions are tested with real physical consequences.

ThinkMorph enables true multimodal chain-of-thought reasoning where text and vision complement each other, achieving emergent intelligence and strong generalization with just 24K training samples.

We contribute EMMA (Enhanced MultiModal reAsoning), a benchmark targeting organic multimodal reasoning across mathematics, physics, chemistry, and coding. SOTA models struggle big time.

We propose Speculative Reward Models that dramatically improve LLM decision-making efficiency while maintaining high accuracy.

A comprehensive survey of using LLMs as evaluators: how to build reliable LLM judges, how to evaluate them, and where they are applied.

We propose CMR Scaling Law to predict critical mixture ratios for continual pre-training, enabling efficient domain adaptation while preventing catastrophic forgetting.