I am studying how to control the way language models use context.I am
Broad research interests
As language models become more capable, it matters more that we can understand and correct their behaviour so that it stays aligned with human values. This is what draws me to alignment, which I work on with a focus on AI safety in Prof. Timothy Baldwin's group at MBZUAI. During my summer visit to Prof. Isabelle Augenstein's group, we found that a short optimized suffix in the input can steer whether a model follows the given context. I want to understand how such behaviours are represented in a model's latent space, and whether mechanistic control and causal representation learning can give us reliable handles on them.
Keyword-based jailbreak evaluation makes steering against the refusal direction look stronger than it is. I tested whether SAE features could reveal a harmful subspace for more actionable steering. The negative results led me to realize that we cannot construct such a subspace, since aligned models only produce refusal outputs, not actionable harmful ones. I then came across GCG as a more principled method, since it looks at the output and tries to maximize it, but compared with refusal-direction steering, it did not work well enough.
Removing the refusal direction jailbreaks a model well enough but at a cost: it must be done inside the model at every generation step. I asked whether an optimized suffix could get the same harmful answers from the input alone. My empirical results showed that it fell well short, while rewriting the request with persuasion strategies came closer. I carried the same tools, GCG suffixes and judge-based scoring, into this paper on context following, where the suffix does not have to fight safety training, and it turned out to work pretty well.
Ho Chi Minh City University of ScienceB.Sc. in Computer Science
VinBigDataAI Engineer Resident
VinBrainAI Engineer Intern
VinUniversityResearch Assistant
FPT Software AI CenterAI Research Resident
MBZUAIM.Sc. in Natural Language Processing
University of Copenhagen, CopeNLUVisiting Student
20182020202220242026
Grants and awards
2026DDSA Visit Grant, Danish Data Science Academy
2024Diversity and Inclusion Award, ACL 2024
2022VINIF Scholarship, Vingroup Innovation Foundation
Unsorted misc
Outside research, I love walking in nature 🌿, travelling ✈️, and soccer ⚽. My favorite player is Toni Kroos, who showed me that simplicity at the highest level is an art.
