Visiting Researcher at Google DeepMind

I study the internal representations of large language models — how pretraining and context shape them, and how they connect to model behavior — through the combined lens of mechanistic interpretability and computational neuroscience. I am now focused on agentic behavior: understanding how agents represent and use information, and giving them strategies for efficient behavior and long-term planning.

Some examples of my research include: investigating the structure of representations in large language models [19] [18] [16] [10]; studying shared representations across biological and artificial neural networks [17] [14]; and exploring how the brain and artificial neural networks build representations of natural language [12] [13] [9].

Previously, I was a Postdoctoral Associate at MIT, where I also completed my Ph.D. in Computational Neuroscience in the lab of Evelina Fedorenko. There, I studied the neural basis of language comprehension using neural recordings, computational models, and LLMs. Prior to that, I worked with Mehrdad Jazayeri on timing and motor variability, and with Wilsaan Joiner on sensorimotor learning. I am originally from Behbahan, a small town in southwest Iran.