Arman Cohan, assistant professor of computer science at Yale, is a two-time recipient of the Roberts Innovation Award from Yale Engineering and Yale Ventures, which supports faculty-led projects that bring breakthrough innovations closer to real-world impact.

His most recent project, SciMentor, is an AI-powered platform designed to serve as a virtual research advisor for early-stage researchers. Built on an advanced LLM-Agent framework, SciMentor offers adaptive, domain-specific mentorship—helping users navigate challenges like limited advisor access, literature overload, and early hypothesis testing.

Cohan was also recently named a recipient of the Google Research Scholar Award, an award recognizing early-career faculty advancing bold and impactful research. His awarded project, Advancing Language Agents for Reasoning-Intensive Scientific Discovery, investigates how AI agents can support the complex reasoning demands of cutting-edge scientific inquiry.

We asked him three questions about how industry recognition, interdisciplinary collaboration, and specialization in AI systems are shaping his work at Yale.
 

Your work sits at the intersection of AI, education, and scientific research. What sets your approach apart from others in the field, particularly when it comes to how researchers interact with AI tools?

I’d say the main differentiator—at least in what we’re currently exploring in the SciMentor project—is the educational aspect. We’re focused on building systems that support researchers in their learning, rather than trying to automate the entire research workflow, which is what some others in the field are aiming for. Our approach is twofold. First, we want to help researchers learn more effectively by collaborating and interacting with AI assistants that are specialized in their domain. Second, on a more technical level, most existing AI agents are quite general—they're capable across many tasks but not deeply specialized. Our goal is to develop technologies that adapt to specific domains in the simultaneously niche and vast world of research opportunities. So, for example, if a researcher is working in a particular subfield of machine learning, we aim to adapt our models to support that domain’s unique needs.

What opportunities do you see for Yale researchers to engage with scientific discovery using AI agents?

AI has a lot to offer researchers today, and especially within the Yale community. One thing we're really interested in is collaborating with domain experts from different departments across campus. One of Yale’s strengths is the breadth and depth of expertise across so many fields, which is incredibly valuable for this kind of work. Our approach is to develop prototypes of our AI methods and share them with researchers to get their feedback. That helps us understand how to make these tools more useful—whether it's to support their scientific discovery process, streamline their research workflows, or even enhance mentorship opportunities.

You recently won a Google Research Scholar Award. How important is this kind of external validation from private industry for the work you do, particularly in computer science?

External validation from industry is encouraging, especially in a fast-moving area like LLMs and LLM agents. These days the frontier research in language modeling and AI is mostly happening in industry. That’s largely due to the enormous computational and engineering resources they have access to. For example, training models with capabilities similar to ChatGPT or Claude requires millions of GPU hours and large teams of highly skilled engineers and scientists, which simply aren’t available in academia at that scale.  But academia plays a critical role in shaping the field’s direction: it provides the space to ask foundational questions, develop new conceptual frameworks, explore new directions, and challenge assumptions that industry may take for granted. Many of the core ideas behind today’s systems originated in academic settings. So, when industry partners recognize and support academic work, it affirms that this kind of long-term, curiosity-driven research continues to have real-world relevance and impact.