Suchir I trace how abstractions emerge across training — the phases of learning, and how data and optimisation shape representation formation. I also maintain open tooling so these experiments are reproducible for everyone.
Research
Research
Cognitively-inspired multilingual AI and the learning dynamics of small language models.
Overview
Research Orientation
I study how neural networks acquire, organise, and transfer knowledge — using language models as a controlled scientific instrument to uncover the mechanisms of learning and representation formation.
My work is guided by a single central question: how do learning systems acquire, organise, and converge upon structured knowledge? Language is the ideal testbed, because it offers rich structure — syntax, semantics, morphology — alongside decades of developmental theory, cross-linguistic variation, and behavioural benchmarks against which learning can be measured.
I studied Computer Science and Linguistics at the University of Cambridge, completing both my undergraduate and master's degrees at Gonville & Caius College, and I am now supervised by Professor Paula Buttery. Over time my interests shifted away from formal descriptions of language toward questions of acquisition: how structure arises without explicit instruction, how learning proceeds unevenly, and how competence is gradually assembled rather than suddenly attained.
My core instrument is the small, human-scale language model (BabyLM) — a deliberately constrained system whose limits make learning visible and measurable. Everything in my research radiates from this core. Much of machine learning is organised around endpoints — benchmarks, leaderboards, final accuracy — and my work is motivated by a discomfort with that framing.
Small language models are not merely scaled-down versions of larger ones. Their value lies less in what they achieve than in what they make intelligible.
I therefore approach language models as epistemic artefacts, focusing on learning dynamics rather than end states, and drawing on tools from psychometrics and developmental psychology — such as item response theory.
The long-term vision is a scientific framework for how structured knowledge emerges, is organised, and becomes transferable in learning systems — combining learning dynamics, representation theory, mergeability, and computational cognitive science. Language models are the focus; the questions concern representation learning broadly.
Directions
Three Strands
Three connected threads run through my doctoral work, from tokenization and learning dynamics to cognitively-inspired evaluation and theoretical linguistics.
Language Models: Learning Dynamics and Tokenization
My research is concerned with building data-efficient small Language Models. While industry-led efforts have built competitive LLMs that have fundamentally shifted the job of the contemporary Natural Language Processing (academic) researcher in various ways, there are still several fundamental, open questions to address that are not obviously ancillary to those pursued commercial AI research labs.
For small LMs, tight parameter budgets make each decision critical, yet researchers still lack systematic, scientific ways to test and refine new ideas. We introduce PicoLM, a lightweight, modular framework that enables systematic, hypothesis-driven research for small and medium-scale language model development. Pico consists of two libraries that together provide a practical sandbox where researchers can make targeted changes to a model's architecture or training procedures and directly observe their effects on the model's behaviour. To support reproducible experimentation, we also release a suite of baseline models, trained under standardised conditions and open-sourced for the community. Check out the YouTube Video put together by Zeb Goriely: Introducing PicoLM | YouTube.
Developmental Interpretability (DevInterp) and Language Model Learning Dynamics Research are fundamental to developing small LMs. Small LMs are crucial for low-resourced and democratised Language Modelling, and should still be a serious focus of hypothesis-driven scientific exploration.
Cognitively-Inspired AI: Linguistic Interpretability and Multilinguality
The second strand of my research focuses on cognitively-inspired AI. The scientific study of the human capacity for language demands a multi-model approach that combines a characterisation of the universal and language- or speaker-specific substance of grammars and the inductive biases that support their emergence in language acquisition, and how these conditions interact with domain-general cognition. This is why I am, in part, drawn to the work of the BabyLM Workshop.
In joint work, we introduce ByteSpan, a dynamic tokenisation scheme that groups predictable bytes rather than pooling their representations.
Linguistics and Cognitive Science
I also actively pursue research interests in Linguistics and Cognitive Science. My work aims to cross-cut modern Generativist approaches – specifically, neo-emergentist approaches pursued by several Cambridge-based theoretical linguists.
The Science · Fundamental
Language Model Science
Four connected threads make up the fundamental core of my work — for academics, cognitive scientists, and computer scientists. Each treats small language models as a scientific instrument for studying how structured knowledge is learned. Tap any card to learn more and see representative papers.
I treat input representation as central to what a model can learn, designing information-driven ways of splitting text so that learning is more efficient, robust, and cross-lingual.
I look inside small models as they train to understand how linguistic structure emerges over development, drawing on linguistics and cognitive science.
I study representation convergence — when independently trained models arrive at compatible representations — and use mergeability as a window onto the geometry of what has been learned.
🧠
Cross-cutting interest · Cognitive science & language acquisition
This is neither purely fundamental nor applied — it is an interest that informs the questions I ask. Decades of linguistic and cognitive theory give precise hypotheses about how humans acquire language. I use cognitive science as an independent source of hypotheses, and build computational models that put those theories to the test.
Applied · Impact
Application areas
The fundamental science radiates outward into three areas of real-world impact, each speaking to a different audience. Tap any card to learn more.
Insights from representation convergence and mergeability translate directly into practical systems: small models that can be composed and merged, and agentic pipelines built from efficient small LMs. Relevant to founders, industry teams, and ML researchers looking to build and collaborate.
I develop multilingual models and benchmarks that reach beyond English to under-served languages, making language technology fairer and more culturally adaptive across the world.
I work on human-centred applications of AI that support learning — including teacher-style demonstrations for language learning, culturally grounded evaluation, and AI for young people. This ties to public engagement through Per Capita Media, and speaks to social scientists, educators, and policymakers.