Suchir Writing
Writing & Essays
Essays on research, journalism, and academic life, from minds to machines.
Essays & Reflections
Essays
Beetle: a framework for controlled bilingual & multilingual pretraining
The open-source Beetle release — 114 controlled models, and why controlled pretraining is worth the effort.
PersonalYear One of a Cambridge PhD: Eight Papers, One Framework, and a Lot of Coffee
Taking stock after the first year of a PhD — the output, the collaborators, and what I would do differently.
AcademicPhonology at Home: Reflections on Hosting OCP23 at Caius
Hosting the Old-World Conference in Phonology at Caius — and what phonology has to teach AI.
SeminarsNLIP Seminars Michaelmas 2024
Personal takeaways from a term of NLIP talks across cognition, linguistics, and language models.
The Papers
On my research
Short write-ups of individual papers, with references in the margins.
Teacher Demonstrations in a BabyLM's Zone of Proximal Development
How a slightly-stronger "teacher" improves multi-turn learning in a small model. Outstanding Paper, EMNLP 2025.
Paper · AwardLooking to Learn: Token-wise Dynamic Gating for Vision-Language Modelling
Token-wise gating for low-resource vision-language models. Outstanding Paper, EMNLP 2025.
PaperByteSpan: Information-Driven Subword Tokenisation
Tokenisation that groups predictable bytes instead of pooling their representations.
Paper · EMNLP 2026LangMAP: A Language-Adaptive Approach to Tokenization
Adapting tokenisation to the language rather than committing to one fixed scheme.
PaperBabyBabelLM: A Multilingual Developmentally-Plausible Benchmark
A multilingual pretraining benchmark built to a child's scale of language input.
Paper · EMNLP 2026Cross-Lingual Alignment Without Joint Training
Do monolingual models converge on universal representations with no joint training?
Paper · FrameworkPico: Hypothesis-Driven Small Language Model Research
A modular sandbox and baseline suite for controlled small-model experiments.
PaperWhat's the Best Sequence Length for BabyLM?
How the training sequence length shapes what a human-scale model learns.
PaperBLiSS: Bilingual Learner Competence in Small Models
Evaluating second-language learner competence in small language models.
PaperLess is More: Cognitively-Plausible Curricula for Cross-Lingual Small Models
Curriculum-learning strategies for pre-training multilingual BabyLMs.
PaperMeasuring Grammatical Diversity from Small Corpora
Annotation-invariant measures of grammatical diversity from limited data.
Paper · BlackboxNLP 2026Development of Linear Truth Encodings in Language Models
A replication study of how linear "truth" directions emerge during training.
Events I organise
Workshops & symposia
The BabyVLM Workshop: Developmentally Plausible Multimodal Systems
Co-organising a NeurIPS 2026 workshop on sample-efficient, infant-scale multimodal learning.
Event · EMNLP 2026BabyLM: Sample-Efficient Pretraining on a Developmentally Plausible Corpus
The shared task and workshop for human-scale pretraining — now with a multilingual track.
EventCambridge Language Sciences Symposium 2025
Organising the poster session at Cambridge's cross-disciplinary language-sciences symposium.
EventThe Human in the Loop: Human–AI Collaboration in Language Assessment
A Cambridge symposium on the reliability, fairness and transparency of AI in language assessment.