Nguyen Minh Quang
PhD student in Computer Science · Singapore Management University
Open to internships
About
I'm a PhD student in Computer Science at Singapore Management University, advised by Prof. Hady W. Lauw. My research asks how a model can keep improving by itself: generating and verifying its own training tasks, knowing where it is uncertain, and remembering more at lower cost.
My work spans reinforcement learning, diffusion models, and efficient language-model architectures.
I'm looking for a research or engineering internship in machine learning. If you have a team working on generative models, RL, or language models, I'd love to chat.
Selected research projects
- Language modeling with 30–86 bytes of context memory per token A shared context store using 30–86 bytes per token across 155M–963M models. Latest results cover language modeling, scaling, and serving capacity; planted-copy tests distinguish strong nearby recall from unresolved far retrieval. Blog
- Self-challenging language models: endless verifiable problems at the edge of ability A pretrained language model plays both task writer and solver, using executable verification instead of a task dataset or judge model. Task generation improves in early runs; gains on external benchmarks have not yet been demonstrated. Blog
- Robust policies from offline data Testing whether the output spread of one diffusion model can replace ensemble disagreement as an uncertainty signal for offline RL. One-step results are promising; policy-level benefits remain unproven. Blog
- Learning to summarize from raw text alone Reinforcement learning for abstractive summarization with no human labels: a model is rewarded when its summary lets the original text be rebuilt, and a frozen anchor reduces reward-driven drift. Blog
Publications
- Per-Starting-Point Plausibility and Diversity Bounds for Score-Based Diffusion Models NeurIPS 2026 Non-asymptotic analysis of the unguided reverse diffusion SDE from a fixed starting point: an upper bound on distance to the data distribution and a lower bound on distinguishability between starts. Paper will be available soonBlog
- Augmenting Decision with Hypothesis in Reinforcement Learning ICML 2024 ALH: give a value-based agent a learned hypothesis about its environment, to fix slow early exploitation and sensitivity to a biased critic. Beats strong baselines on MuJoCo. Paper Code Blog
- Zero-cost Transition to Multi-Document Processing in Summarization with Multi-Channel Attention ECML PKDD 2024 · undergraduate work at VNU Hanoi, before the PhD Reuse single-document BART weights for multi-document summarization with linear (not quadratic) cost in the number of documents, no re-training needed. Paper Code Blog
* shared first authors