PhD Candidate · University of Cambridge
I study how AI systems learn — treating small, human-scale language models as scientific instruments.
PhD candidate in Computer Science at the University of Cambridge (Gonville & Caius College), working across machine learning, cognitive science and linguistics to understand how language models acquire capabilities and develop internal representations during training.
Research at a glance
I study learning as it unfolds through training. Rather than only measuring what a finished model can do, I use small, human-scale language models as controlled experiments — tractable enough to inspect directly, yet rich enough to test hypotheses drawn from human language acquisition and cognition. Four connected questions run through the work.
How do capabilities and internal representations emerge during training — and can we monitor that development before standard evaluations detect it? This connects interpretability to AI safety.
Learning dynamics & interpretabilityHumans learn language from far less data than today's models. Which architectures and training strategies let a model acquire real capabilities from child-sized data budgets? This drives the BabyLM and BabyVLM challenges.
Human-scale modellingMost people use more than one language, yet models of language processing are largely monolingual. By controlling what languages a model sees, and when, I test how linguistic experience shapes what it learns and generalises.
Beetle model suiteTokenization is usually a fixed preprocessing step. I treat it as a learned resource-allocation policy: can a model decide how much computation to spend on different parts of its input?
Efficient architecturesSelected projects
Two open-source projects anchor my research programme — one a suite of controlled models, the other the framework that trains and studies them.
Beetle · Bilingual model suite
The problem. Computational models of human language processing have focused heavily on monolingual speakers, even though most people in the world use more than one language.
What I build. Beetle is a framework and a collection of 330 small language models trained under controlled bilingual and monolingual settings — matched populations that differ in a specific aspect of their learning history.
Why it matters. This experimental control lets us test hypotheses about bilingualism — code-switching, language attrition, heritage learning — that are difficult to study in larger, less transparent systems.
PicoLM · Learning-dynamics framework
The problem. Studying how capabilities emerge requires training many small models and inspecting them checkpoint by checkpoint — infrastructure most researchers rebuild from scratch.
What I build. PicoLM is an open framework for pretraining, interpreting and tokenizing small language models, designed for hypothesis-driven experiments on learning dynamics.
Why it matters. It makes small-model research reproducible and comparable, giving a common foundation for the BabyLM community and my own work.
Recent news
Recognition
Recognised at the BabyLM Workshop in Suzhou for work on sample-efficient, human-scale language modelling.
Recent work selected for oral presentation at EMNLP 2026.
Working with me
I am looking for motivated MPhil students interested in developing and understanding language models. Projects are based in the Cambridge Natural Language and Information Processing (NLIP) group, with opportunities to collaborate across Computer Science, Linguistics and Cognitive Science.
About
I am a PhD candidate in Computer Science at the University of Cambridge (Gonville & Caius College), supervised by Professor Paula Buttery. My research sits at the intersection of machine learning, computational linguistics, cognitive science and AI safety.
I previously completed a BA and MEng in Computer Science and Linguistics at Cambridge, earning a Starred First (Class I with Distinction) and a Distinction. My master's work drew me into the BabyLM Challenge, and from there into the broader question of how models acquire their capabilities.
I co-organise the BabyLM and BabyVLM research challenges, lecture and supervise in Cambridge, and organise the weekly Cambridge NLIP seminars, open to the public.
Get in touch
I am always happy to hear from prospective students, collaborators and anyone curious about small language models.
Suchir