Research

I am interested in building AI agents that learn from feedback and support scientific discovery. My work spans reinforcement learning for generative models, agent workflows for biomedical machine learning, reasoning in large language models, and fairness in speech generation. Click a topic to read more.
Project Co-Lead · Northwestern University

Reinforcement Learning for Discrete Generative Models

Fine-tuning discrete flow-matching models with reinforcement learning, applied to regulatory DNA and protein design.

  • +10.6% DNA predicted activity
  • +15.4% protein pred-ddG
Project Lead · Northwestern University

AI Agents for Biomedical Discovery

cfDNA-Scientist, a self-evolving multi-agent system that proposes, implements, and evaluates biomedical features from cell-free DNA data.

  • AUC 0.916 vs. 0.831 baseline
  • Breast cancer detection
First Author · Inventec AI Center

Reasoning in Large Language Models

A dual-process scaffold reasoning framework that makes LLMs better at debugging code.

  • 88.91% pass rate on DebugBench
Co-Author · INTERSPEECH 2026

Bias in Speech Generation

How multiple cues in natural-language instructions combine to form gender bias in instruction-following text-to-speech.

  • Accepted to INTERSPEECH 2026

See also my projects and publications.

Project Co-Lead · Northwestern University

Reinforcement Learning for Discrete Generative Models

Discrete flow-matching models generate discrete sequences such as DNA or proteins by gradually transforming noise into structured data. I study how to optimize these models as policies with reinforcement learning, so they can be steered toward outcomes defined by a reward signal.

As project co-lead, I co-developed DoMinO, a reinforcement learning framework that formulates discrete flow-matching inference as an inner Markov decision process. This enables exact trajectory log-likelihood optimization. We fine-tune with REINFORCE- and PPO-based methods, using cross-entropy and generalized KL regularization to mitigate reward hacking and preserve sequence naturalness.

On regulatory DNA and protein design, DoMinO outperforms prior state-of-the-art baselines by 10.6% on DNA predicted activity and 15.4% on protein pred-ddG.

DoMinO project Paper arXiv

Project Lead · Northwestern University

AI Agents for Biomedical Discovery

Finding informative features in biomedical data usually takes many rounds of expert trial and error. I led cfDNA-Scientist, a self-evolving multi-agent system that autonomously proposes, implements, and evaluates feature-extraction code for cell-free DNA (cfDNA) data in executable research environments. The agents coordinate several stages:

  1. Hypothesis generation: propose candidate biomedical features.
  2. Feature implementation: turn each hypothesis into working code.
  3. Verification: check that the implementation is correct.
  4. Critique: evaluate the feature and point out weaknesses.
  5. Curation: keep the most useful features for downstream modeling.

A preliminary literature-search workflow first identifies established cfDNA fragmentomic features, builds a feature catalog, and validates baseline performance. On breast cancer detection, cfDNA-Scientist reaches a cross-validated AUC of 0.916, compared with 0.831 for the baseline under the same evaluation protocol.

cfDNA-Scientist project

First Author · Inventec AI Center

Reasoning in Large Language Models

I am interested in how to structure the reasoning of large language models so they solve hard, multi-step problems more reliably. During my internship at the Inventec Artificial Intelligence Center, I proposed a dual-process scaffold reasoning framework for LLM code debugging. It reaches an 88.91% pass rate on DebugBench, outperforming prior prompting methods.
Paper arXiv

Co-Author · INTERSPEECH 2026

Bias in Speech Generation

Instruction-following text-to-speech (TTS) systems let users describe the voice they want in natural language. I contributed to an analysis of how multiple cues in these instructions bind together to form gender bias in the generated speech.
Paper arXiv