Year One of a Cambridge PhD: Eight Papers, One Framework, and a Lot of Coffee

Eighteen months ago I started a PhD. I had just submitted my MEng thesis — a paper on cognitively-plausible training curricula for multilingual BabyLMs, eventually accepted at the BabyLM Workshop at EMNLP 2024 in Miami — and I had the slightly vertiginous feeling of having finished one thing and not quite started the next.

This post is an attempt to take stock. Not the kind of stock-taking you do for a progress report or a grant application, but the kind where you try to be honest about what you've learned, what surprised you, and what you'd tell yourself if you could go back to October 2024.

The output

By the end of my first year, eight papers had been accepted: four at the BabyLM Workshop at EMNLP in Suzhou, two at CogInterp at NeurIPS in San Diego, one in the EMNLP Systems Demonstrations track (the PicoLM paper), and one in the Multilingual Representation Learning workshop. Two oral papers at OCP23 in January 2026. A Computational Linguistics journal paper at ACL in Vienna in July. ByteSpan at the ICML Tokenisation Workshop in Vancouver.

Listing these feels simultaneously impressive and somewhat embarrassing. Impressive because it represents a lot of work, a lot of travel, and a lot of late nights. Embarrassing because I know how much of it was collaborative — Richard, Zeb, Paula, Fermín, Bianca, Yuan, David, all the ALTA people — and I sometimes worry that the convention of listing the first author's name gives a misleading picture of how research actually gets done.

The PicoLM framework was the centrepiece. We released it in March 2025, and the reception was better than I had hoped. Building a research infrastructure that other people actually use is a different kind of satisfaction from writing a paper. Papers are arguments; frameworks are invitations. The fact that two UROP students spent the summer working with it, that collaborators from KAIST and elsewhere have engaged with it, that it's now the basis for several BabyLM 2025 submissions — that feels like something that might matter beyond a single paper cycle.

What I learned about research

The most important thing I learned in my first year is that the interesting questions are usually not the ones you started with. The phonology work with Fermín was not in my original research plan. ByteSpan emerged from a conversation about tokenisation that started as a brief detour from something else. The NeurIPS CogInterp position paper began as a rant in a supervision that Paula diplomatically suggested I write up properly.

This is not a peculiarity of my work; it's how research works. The research programme you write in your proposal is a useful fiction — it tells the funder (or the admissions committee) that you have thought carefully about what you're doing, which is true. But it is not a map of where you will actually go.

The second thing I learned is that good collaborators are everything. I have been extraordinarily lucky in this regard. Paula gives me rope and intellectual companionship in roughly equal measure. Fermín brings a depth of theoretical knowledge that makes me feel perpetually, productively out of my depth. Richard has built PicoLM with a combination of technical rigor and strategic clarity that I aspire to. Zeb's ability to communicate complex ideas simply — the PicoLM YouTube video is a masterclass — is something I am still trying to learn.

What surprised me

The teaching. I didn't expect to find it as rewarding as I do. Being a Teaching Assistant for CST IA Machine Learning & Real World Data, supervising the BabyLM workshop group, guest-lecturing on Language Model Evaluation: these are time-consuming, often humbling (students ask questions I can't always answer), and genuinely one of the best parts of the job. There is something about explaining a concept well that sharpens your own understanding in a way that writing about it doesn't.

The organisational work also surprised me. Running the NLIP seminars, sitting on the OCP23 organising committee, co-chairing the Language Sciences poster session: none of this was in the plan, and all of it was valuable. Not for the reasons people usually cite ("good for your CV") but because it puts you in contact with work outside your immediate area, which is where the unexpected connections come from.

What I'd change

I would protect more time for thinking. The first year of a PhD is, structurally, very similar to a structured procrastination engine: there is always a deadline, always a workshop to submit to, always a meeting to prepare for. The deep, uninterrupted thinking that produces the best ideas gets squeezed by the busyness of being a productive early-career researcher.

I am trying to build better habits in year two. More writing in the mornings before email. More walks. More seminars outside my area — I've been going to the Cavendish Theoretical Physics seminars when I can, partly from curiosity and partly because physicists think about emergence and scale in ways that are relevant to what I do, even when the vocabulary is completely different.

A note on independence

There is a particular kind of freedom in a PhD that doesn't exist elsewhere in the academic career structure. You are not yet accountable for a lab, a grant, a teaching load. The scope of what counts as "your research" is broader than it will ever be again. The penalty for following an interesting tangent is low.

I am trying to use that freedom well. The phonology work, the journalism, the framework building — these are all, in different ways, expressions of a conviction that the most interesting intellectual work happens at the intersections, and that the intersections are most easily navigated when you haven't yet been fully captured by a single disciplinary identity.

Year two starts now. More to come.