Research

I build and study small language models trained on developmentally plausible data, in and around the BabyLM Challenge. Under tight budgets each design choice — curriculum, sequence length, supervision signal — becomes measurable, so these models double as scientific instruments for the mechanisms of acquisition. Representative papers: Teacher Demonstrations (ZPD), Best Sequence Length for BabyLM, Less is More, BabyLM Turns 4 (2026 shared task).

I trace how abstractions emerge across training and look inside small models as they develop, drawing on linguistics and cognitive science. I maintain open tooling — PicoLM — so these developmental-interpretability experiments are reproducible for the wider community. Representative papers: Pico, Theoretical Linguistics & Causal Abstraction, Claude reinvestigates herself, Maximising Minimal Means.

I study reinforcement-learning post-training — what constrains exploration under algorithms like GRPO — and alignment objectives inspired by how teachers scaffold learners. Ongoing work includes From Entropy to Exploration: What Limits Post-Training under GRPO? and work on pedagogical alignment. Representative papers: Pedagogical Alignment of LLMs, Protocol Competence of Small LMs.

My flagship project, Beetle, uses controlled bilingual pretraining to isolate how a second language is acquired and how cross-lingual transfer arises. This connects to multilingual benchmarks and curricula that reach beyond high-resource English into under-served languages. Representative papers: Beetle, BabyBabelLM, BLiSS, L1 Influence in L2 Models, Meta-Pretraining for Cross-Lingual NER, Multimodal Grounding across Languages & Cultures.

I design information-driven ways of splitting text so that learning is more efficient, robust and cross-lingual — grouping predictable bytes dynamically, and studying language-adaptive tokenisation that lets the input representation shift with the language being modelled. Representative papers: ByteSpan, LangMAP, Is Tokenizer Cognitive Plausibility just Fertility?.

Through the xBLiMPs grant and related community work, I develop minimal-pair and grammaticality benchmarks that extend equitable linguistic evaluation to low-resource languages — the infrastructure needed to make fairer multilingual claims credible. Representative papers: BabyBabelLM, Measuring Grammatical Diversity, BLiSS, Phoneme Frequencies across Languages, Phonemes: Chance, Costs & Tiers, Diachronic Phoneme Frequencies, Convergent Phoneme Surprisal.

I study cultural knowledge in pretrained language models — including code-switching and cultural localisation — and argue that the instability of cultural responses is itself informative. Ongoing work includes Repertoires, Not Scores: Instability as Signal in Cultural Evaluation of LLMs. Representative paper: Repertoires, Not Scores.

I study representation convergence — when independently trained models arrive at compatible representations — and use model merging as a window onto the geometry of what has been learned. Ongoing work includes Mergeability is Constrained by Initialisation, Similarity is not Mergeability, and Representations Bound the Generalizability of Model Merging. Representative papers: Merged Monolingual Goldfish Models, Cross-Lingual Universal Conceptual Representations, Cross-Lingual Alignment Without Joint Training, Linguistic Universals via Sparse Autoencoders.

I study how language-model representations align with the brain across linguistic domains, and use mechanistic interpretability to locate structure inside models — syntactic-agreement units and linear encodings of truth among them. Ongoing work includes Distinct Patterns of Brain Alignment of LMs across Linguistic Domains. Representative papers: Theoretical Linguistics & Mechanistic Interpretability, Claude reinvestigates herself, Linear Truth Encodings, Linguistics in the Age of Language Models.

Evaluation is where scientific claims are won or lost. I work on whether small language models can serve as judges, and on the governance and evidence standards for trustworthy compact AI. Ongoing work (NeurIPS 2026 workshops) includes Can We Trust Small Language Models as Judges?, Towards Trustworthy Compact Judges, and Governing the Evidence Standard for Trustworthy Compact AI. Representative papers: Pedagogical Alignment of LLMs, Repertoires, Not Scores.

I work on agentic uses of small LMs that are cheap enough to run on-device, including execution-aware routing and the protocol competence needed for interactive language games. Ongoing work includes Execution-Aware Routing for On-Device SLM Agents and studies of protocol competence in interactive settings. Representative paper: Protocol Competence of Small LMs.

I work on AI safety and governance, with a focus on vulnerable audiences and the gap between recognising risk and acting on it. Ongoing work includes Models Recognise Child Audiences, But Child Safety Does Not Follow and related position papers. Representative papers: Repertoires, Not Scores, Pedagogical Alignment of LLMs.

Insights from representation convergence, mergeability and RL post-training translate directly into practical systems: small models that can be composed and merged, and execution-aware agentic pipelines built from efficient small LMs. Relevant to founders, industry teams, and ML researchers looking to build and collaborate.

I develop multilingual models, bilingual pretraining, and evaluation that reaches beyond English to under-served languages — including equitable benchmarks through the xBLiMPs grant — making language technology fairer and more culturally adaptive across the world.

I work on human-centred applications of AI that support learning — teacher-style demonstrations for language learning, culturally grounded evaluation, child-safety, and AI for young people. This ties to public engagement through Per Capita Media, and speaks to social scientists, educators, and policymakers.