I study how cognitive maps form in the human brain and how their neural representations change with learning. Previous work shows that 2-D, map-like representations can emerge in regions such as the entorhinal cortex and hippocampus, while the prefrontal cortex tends to compress them into 1-D. In my PhD I combine behavioral experiments, fMRI, EEG, computational models, and RNNs to ask how neural geometry changes during learning, what drives the hippocampus and PFC toward different representations, and how these representations support flexible inference and generalization. 我研究人类认知地图如何在大脑中形成，以及其神经表征如何随学习而变化。已有研究表明，内嗅皮层与海马等脑区可以学到二维的、类地图的表征，而前额叶皮层则倾向于将其压缩为一维。博士期间，我结合行为实验、fMRI、EEG、计算模型与 RNN，探究学习过程中神经几何如何变化、是什么驱动海马与前额叶形成不同的表征，以及这些表征如何支持灵活的推理与泛化。

LLMs like Centaur show remarkable performance in modeling human behavior, but their heavy data requirements make them hard to fine-tune for predicting and explaining behavior — and they may take shortcuts that capture choice preferences rather than the underlying task structure. To address this, we train a Transformer on the large-scale behavioral dataset from Poldrack Lab's 2019 Nature Communications paper, in which hundreds of participants completed a comprehensive battery of cognitive tasks. We believe a dataset this comprehensive opens the door to a model that captures the structure of human intelligence. 像 Centaur 这样的大语言模型在刻画人类行为上表现惊人，但庞大的数据需求让它们难以微调来预测和解释行为，而且可能走捷径，只捕捉选择偏好而非任务的内在结构。为此，我们基于 Poldrack 实验室 2019 年 Nature Communications 论文的大规模行为数据集训练一个 Transformer——该数据集包含数百名被试完成的一整套综合认知任务。我们相信，如此全面的数据集让我们有可能构建出真正捕捉人类智能结构的模型。

Wu et al. (2024) show that human altruistic behavior is driven by a combination of seven socioeconomic motives — a "motive cocktail." With Xiaoyan, I use EEG and MEG recorded while participants perform the same task to decode the temporal and neural dynamics supporting each type of motive. Wu 等人（2024）发现，人类的利他行为由七种社会经济动机的组合所驱动，称为"动机鸡尾酒"。我与吴小燕合作，利用被试执行同一任务时记录的 EEG 与 MEG，解码支持每一种动机的时间与神经动态。

Artificial neural networks predict human behavior far better than classical cognitive models, but their many parameters make the underlying algorithms hard to interpret. Using HybridRNN — a framework from DeepMind — we combine the accuracy of ANNs with the interpretability of cognitive models to study human learning. With Maria, I apply HybridRNN to the computations behind flexible adaptation in reversal learning, and use dynamical-systems analysis to understand its internal representations. 相比经典认知模型，人工神经网络能更准确地预测人类行为，但其庞大的参数量让背后的算法难以解释。借助 DeepMind 提出的 HybridRNN 框架，我们把神经网络的精度与认知模型的可解释性结合起来研究人类学习。我与 Maria 合作，用 HybridRNN 刻画反转学习中灵活适应背后的计算过程，并通过动力系统分析理解其内部表征。

At EBKernel, I work on a next-generation vision-language navigation (VLN) system grounded in findings about human cognitive maps. Building on the Tolman-Eichenbaum Machine and Vector-HaSH models, the entorhinal-hippocampal circuit offers an effective framework — a predefined spatial grid backbone plus efficient item-space binding — that lets humans generalize spatial and non-spatial knowledge across environments. Our goal is a brain-inspired VLN system with strong flexibility and generalization. 在 EBKernel，我致力于一套以人类认知地图研究为基础的新一代视觉语言导航（VLN）系统。基于 Tolman-Eichenbaum Machine 与 Vector-HaSH 模型，内嗅-海马环路提供了一个高效框架——预设的空间网格骨架加上高效的"物体-空间"绑定学习——使人类能够在不同环境中泛化空间与非空间知识。我们的目标是打造一个灵活、可泛化的脑启发 VLN 系统。