I'm a PhD student in EECS at MIT where I am fortunate to be advised by Aleksander Mądry. I received my BS in computer science from Stanford, and spent two great years at Robust Intelligence before starting my PhD.
I am interested in how we can develop AI systems that can be safely deployed. Outside of research, I enjoy running, climbing, skiing, tennis and volleyball.
For a full list of papers, see Google Scholar.
Benjamin Cohen-Wang*, Harshay Shah*, Kristian Georgiev*, Aleksander Mądry
Language models may need external information to provide a response to a given query. We would provide this information as context and expect the model to interact with it when responding to the query. But how would we know if the model actually used the context, misinterpreted anything, or made something up? We present ContextCite, a method for attributing statements generated by language models back to specific information provided in-context.
Benjamin Cohen-Wang, Joshua Vendrow, Aleksander Mądry
Pre-training on a large and diverse general-purpose dataset and then fine-tuning on a task-specific dataset can be an effective approach for developing models that are robust to distribution shifts. In practice, this method helps significantly in some cases, but not at all in others. In this work, we characterize the failure modes that pre-training can and cannot address.
We use ContextCite to detect unverified statements and discover poisoned documents.
We present ContextCite, a method for attributing statements generated by language models back to specific information provided in-context.
We explore a simple principle for harnessing pre-training to develop robust models.
We study the robustness benefits of pre-training and characterize failure modes that pre-training can and cannot address.
Attribute (or cite) statements generated by LLMs back to in-context information.
Last updated on March 6, 2025