Giang Nguyen

• Train: First large-scale inherently interpretable language model, Steerling-8B, pretrained on trillions of tokens.
• Explain: Attributing model behavior at the concept, prompt, and input-to-concept levels.
• Control: Controllable generation for topic steering and safety alignment without fine-tuning.

Research on interpretable LLMs for tabular data.

Thesis: Transforming the black-box decision-making of AI models into explain-then-answer processes

Thesis: Overcoming catastrophic forgetting by deep visualization

116-page technical report on how we scaled interpretable AI to 8B parameters. One finding: interpretability improves with scale: the more compute you put in, the more interpretable the model becomes. [pdf]

A scientific video showing the key findings for interpretability scaling: how interpretability scales with compute, what interpretability looks like at 8B parameters, and what this means for building AI you can trust.

Concept-level steering in real time: highlight text, see which internal concepts drove it, steer them, compare outputs side-by-side. [view on X ↗]

My first solo paper. A position paper arguing that feature attribution in generative language models is inherently under-specified: the same attribution method can answer different questions depending on what you are explaining.
[arXiv] [pdf]

The first large-scale language model with interpretability built into the architecture: every generated token traces back to input context, training data, and human-understandable concepts. Featured in TechCrunch.
[blog] [github] [huggingface] [techcrunch]