About Me

I am an LLM Researcher with a Ph.D. in Computer Science from Columbia University, an M.S. from Carnegie Mellon University, and a B.A. from Dartmouth College. My expertise lies in training and evaluating Large Language Models (LLMs) on large-scale, real-world data, with a focus on biomedical and clinical applications. My work aims to bridge the gap between cutting edge research (synthetic data, efficient inference, extending inference compute, etc.) and practical applications.

Research Focus

My research addresses key challenges in long-context language modeling, particularly on noisy datasets. I’ve developed new methods for:

• Dataset curation from real-world data (ClinSum, NAACL ‘21)
• Synthetic data for self-correction (Reference Revision, EMNLP ‘22)
• Offline RL (Summary Calibration, ACL ‘23)
• Knowledge distillation for domain transfer learning (ClinDistill, MLHC ‘23)
• Planning algorithms (SPEER, CoLM ‘24)
• Scaling inference compute (Plan-Guided Best-of-N, ACL ‘23, Chain of Density, EMNLP ‘23)
• Graphical models for representation learning (Latent Meaning Cells, MLHC ‘20)

Current Work

I’m currently focused on reducing the cost of long-context inference. I developed Cold Compress, an open-source library for KV cache compression. Cold Compress implements complex KV cache eviction methods with static computational graphs, which are fully torch compilable. The end result is a toolkit that strikes a balance between simplicity and performance, making it both accessible to all and performant enough for experimentation.

I’m committed to developing models that can effectively handle extended contexts on noisy, real-world data. Please reach out if you’d like to chat about scaling inference-time compute for non-reasoning tasks, making RAG pipelines more friendly to noisy data, or enabling long-context modeling on consumer hardware.