About me: Hello world! You can call me Quan. Currently, I am an undergrad student at HUST, Vietnam, and also an AI Resident at Qualcomm AI Research. I love playing with large language models. I am impressed by the abilities of LLMs, and I am curious about how they work. I want to understand them theoretically.

Updates

• [Dec 2025] I am actively looking for a PhD position. Good luck to me!

• [Jun 2025] I will attend International Conference on Machine Learning 2025. Let’s connect!

• [May 2025] I got 2 papers accepted at ICML 2025. What an achievement for me!

• [Apr 2025] I officialy joined Qualcomm AI Research as an AI Resident.

• [Dec 2024] I achieved an IELTS Academic band score of 7.0!

• [Aug 2024] I joined VinAI Research as an AI Resident.

• [Apr 2024] My first paper was accepted at ICASSP 2024!

Research Interests

Despite their impressive emergent abilities 1, Large Language Models (LLMs) are, at their core, massive statistical engines optimizing a conditional probability distribution. The recent history of our field can be viewed as a sequence of attempts to optimize this probability more effectively. Initially, we relied on prompt engineering to provide hand-crafted conditions to guide generation. This evolved into parameter-efficient methods such as prompt-tuning and prefix-tuning, which learn continuous prompt representations to condition the model. As we reached the limits of static conditioning, the frontier shifted to the post-training phase and test-time compute.

However, recent findings 2 suggest that post-training with reinforcement learning (RL) often acts primarily to \textit{sharpen} the conditional probability of the original LLM. This implies that many of the benefits of RL post-training can be approximated by sampling from a power distribution of the base model distribution 3. In other words, once we scale models sufficiently in both data and parameters, LLMs already possess rich latent abilities, and the challenge becomes how to elicit them reliably. My research therefore aims to

• (i) leverage these intrinsic abilities through robust fine-tuning and test-time compute
• (ii) deepen the scientific understanding of language models by investigating their fundamental statistical properties.