I work on improving a machine’s ability to understand and engage with the world. I’m interested in how society will decide to shape the way we interact with technology, approach problem-solving, and think about creativity.
At Imperial College London, I lead a research group focused on AI interpretability and robustness. I consult on AI deployment, training, fine-tuning, retrieval, and safety. I also work with founders and investors on complex, interdisciplinary problems.
AI models are (seemingly) highly capable, but there are lots of challenges and unanswered questions about their behavior. I’m currently looking into:
Refined control Task performance depends heavily on how prompts are structured. Post-training has advanced models’ instruction-following capabilities, but language is more a bridge than a perfect mirror of thought: some concepts defy precise expression through text, and crafting the right contexts to elicit specific behaviors can be challenging. Furthermore, models remain fragile: small prompt tweaks or adversarial inputs can lead to harmful or erroneous outputs. Research shows that models encode high-level representations of the world. I explore how we can leverage these representations to obtain an additional lever of control on the model’s behavior.
Continual learning Embedding large language models into larger systems with long-term memory and tools for interacting with the world could unlock even more sophisticated task-solving capabilities. Extending context windows is one approach, but it has limits: no window can encompass all past events, nor can we guarantee effective retrieval of events through attention mechanisms. Equally important is the ability to break down complex tasks into structured plans and learn effectively from observations. Enhancing these capabilities is a key area of my research.
Creative problem-solving What will it take to make an AI into a truly useful collaborator? Recent advances in reasoning models are impressive, but these tasks often test for the final answer; the step-by-step reasoning that real-world problems demand is harder to evaluate. I look into how we can build models that can solve complex, open-ended problems—where the process matters just as much as the outcome.
What began with a younger me writing stories about a mom and dad gifting their daughter a cat (an attempt at manifestation?), has evolved into reflections on machine learning and, occasionally, the quirks of life. Visit my Substack if you’re interested to read more.
Prefer to learn about machine learning through videos? Head over to my YouTube channel for deepdives and tutorials on the latest in ML.
See Google Scholar for a full overview of my more formal research publications.
I’d love to hear from you! Whether you have an idea to discuss, a problem to brainstorm, or a project to collaborate on, feel free to reach out. To ensure I read your message, include ‘Building better ML’ in the subject line.
This is a jekyll based resume template made by Ankit Sultana