Nihal Nayak
I’m on the job market. Reach out if you’re interested in hiring me!
I am a Postdoctoral Fellow at Harvard University (SEAS), working with David Alvarez-Melis. I completed my Ph.D. in Computer Science at Brown University, advised by Stephen Bach.
My work introduces scalable, data-centric learning techniques to efficiently supervise and audit frontier AI systems. They span three connected areas:
-
Principled data curation: Efficiently building the next generation of AI systems requires automated methods for curating high-quality training data. My recent work on Bonito focuses on creating synthetic post-training data from unstructured expert documents. Building on that foundation, my work on targeted instruction selection identifies how to select post-training data that predictably improves performance on target tasks.
-
Efficient and flexible models: In collaboration, I discovered boomerang distillation, a simple technique that creates families of LLMs with fine-grained model sizes through model size interpolation while requiring 14–19× fewer training FLOPs than individual distillation.
-
Auditing generalization: I use techniques and concepts from cognitive science to create diagnostics for evaluating AI systems. For example, our concept-binding benchmark shows that vision-language models such as CLIP struggle to associate attributes with the correct objects, failing to reliably distinguish a red cube and a yellow sphere from a yellow cube and a red sphere.
news
| Sep 25, 2026 | Our work on understanding the role of layer patching in model size interpolation was accepted to NeurIPS 2026. |
|---|---|
| Sep 14, 2026 | Invited talk at the University of Pennsylvania, CLunch. |
| Aug 24, 2026 | Our work on ADAPT, which extends boomerang distillation across post-trained model families, was accepted to Findings of EMNLP 2026. |