Research experience

May 2026 - Aug 2026Boston, MA

IBM Research, Core AI Team

Research InternMentor: Akash Srivastava

When does the harness matter for LLM-driven discovery?

Led a controlled comparison of 11 LLM search harnesses, from best-of-N sampling to evolutionary search and coding agents, on math construction, GPU kernel optimization, and nanoGPT training, under a shared task interface, evaluators, and API-cost ledger.
Found that choosing a harness often mattered more than upgrading the model, but the best harness depends on the setup: rankings change with how the budget is measured and which model is used.

Aug 2024 - PresentDurham, NC

DukeNLP, Duke University

PhD StudentAdvisor: Monica Agrawal

How faithfully do LLMs acquire probabilistic knowledge?

Continually pretrained OLMo-2 1B and 7B on millions of notes generated from 650 synthetic Bayesian networks, so learned probabilities can be checked exactly.
Found relational compression: models get individual probabilities about right and rank outcomes correctly, yet understate how far apart outcomes are and how much strong evidence should shift a prediction. Standard metrics miss this.

How do LLMs learn clinical knowledge?

Built MedLingo, a jargon benchmark from real clinical notes, and showed that term–meaning co-occurrence counts in pretraining corpora predict which jargon a model understands; jargon common in clinical notes is often rare in pretraining data.
Models reject disputed medical claims when asked directly but are far likelier to endorse them when the question presupposes them.

Can counting match an LLM?

Introduced FBPR, a Naive Bayes that ranks MedQA diagnoses using concept–diagnosis co-occurrence counts from an LLM's pretraining corpus; it matches that LLM's diagnostic accuracy (OLMo 7B, LLaMA 65B).
The two get largely different questions right, so LLMs are not simply aggregating frequencies, and a simple hybrid of the two beats both.

Interpreting dataset shift in clinical notes

Built a benchmark that induces controlled shifts in real MIMIC-IV notes, and SIReNs, which explains a shift by showing an LLM representative notes from each dataset.
SIReNs nearly matches an oracle on clear-cut binary shifts but struggles with gradual continuous ones. Now extending this to LLM agents and harnesses on multifactorial shifts.

May 2023 - June 2024Los Angeles, CA

Melady Lab, University of Southern California

Research AssistantAdvisor: Yan Liu

GPT4MTS: Prompt-Based Large Language Model for Multimodal Time-Series Forecasting

Built an LLM pipeline that pairs time series with related news summaries and used it to create a GDELT news-coverage dataset for 55 U.S. regions.
Feeding the text to GPT-2 as soft prompts cut MSE by 4% over the best LLM baseline.

TEMPO: Prompt-based Generative Pre-trained Transformer for Time Series Forecasting

Co-developed and ran training experiments for a GPT-2 forecaster that splits each series into trend, seasonal, and residual parts and uses prompts to adapt to distribution shift.
In zero-shot forecasting on unseen datasets, it outperformed baselines evaluated the same way (e.g., 19% lower MAE than PatchTST on ETTm1).

Aug 2022 - June 2024Los Angeles, CA

GLAMOR Lab, University of Southern California

Research AssistantAdvisor: Jesse Thomason

I2I: Initializing Adapters with Improvised Knowledge

Trained CLIP-BART for a method that starts each new task's adapter from a distilled blend of earlier ones.
On five VQA tasks it raised new-task scores 4% relative to independent adapters, while AdapterFusion lowered them 2%: reuse works best as a starting point.