Ya Gao

I am a 3rd-year PhD student at Aalto University School of Science, supervised by Pekka Marttinen. My research focuses on Large Language Models, particularly post-training, LLM Reasoning & Agency, foundation model self-improvement, and applications in healthcare.

Previously, I obtained my Master degree in Machine Learning from Aalto University and Bachelor degree in Computer Science from Sichuan University.

Email  |  CV  |  Google Scholar  |   | 

The recent progress in LLM capabilities has been driven by massive scaling of pre-training data. However, high-quality human data is a finite resource, and relying solely on imitation learning inherently caps model performance at the level of human demonstrators. My research aims to move beyond this ceiling by studying how models can evolve and improve themselves continuously with only limited human supervision.

Specifically, I focus on: (1) utilizing self-generated reasoning trajectories and multi-model interaction to bootstrap performance without extensive human labeling, and (2) analyzing learning dynamics within reasoning traces (e.g., how supervision format, reasoning depth, and data composition affect model behavior) to enable more effective self-improvement. In terms of real-world impact, I am interested in translating these advances into healthcare applications.

We treat knowledge internalization as a reasoning problem and propose a training procedure that introduces knowledge as background stories, enforces its use through multi-hop reasoning, and distills the resulting behavior.

We train LLM agents to internalize knowledge and skills for multiple tasks without relying on ever-expanding prompts or prior demonstrations, through context distillation and efficient use of corrective feedback from humans.

We propose a low-overhead framework that steers LVLMs towards visually-grounded generation by leveraging a single-pass intervention, to address the trade-off between effectiveness and efficiency.

We evaluate the potential of LLMs in analyzing digital phenotyping data to predict changes in depression severity among individuals experiencing major depressive episodes.

We introduce a novel query-guided self-supervised domain adaptation approach for abstractive nursing note summarization. The method uses patient-related clinical queries for guidance, and hence does not need reference summaries for training.

Design and source code from Jon Barron's website.