Suchir I'm Suchir, an AI researcher at the University of Cambridge. I research how AI systems learn and gain capabilities. I contribute to an emerging field known as developmental interpretability, which has applications to Cognitive Science and AI Safety. I am interested in building fully open-source AI systems that are socially and culturally-situated for decision-making, learning and Human-AI collaboration.
I am a PhD candidate in Computer Science, supervised by Professor Paula Buttery, working on the science of language modelling.
Three questions motivate the work. How do representations and capabilities emerge as a model learns? How does the organisation of those representations shape generalisation, composition and transfer? And how can we monitor and evaluate learning systems scientifically, beyond static benchmarks?
I approach them with small, human-scale language models trained under controlled conditions: models small enough to inspect checkpoint by checkpoint, yet rich enough to test hypotheses drawn from language acquisition and cognition. With Catherine Arnett, James Michaelov and Paula Buttery I built Beetle, a suite of 330 open models for studying how language-learning history shapes linguistic representations. As Head of Research of PicoLM, with Richard Diehl Martinez, I help maintain the open framework that trains and analyses them. I co-organise the BabyLM workshop and shared task at EMNLP and BabyVLM at NeurIPS, which ask how much a model can learn from a child-sized data budget. Work from this programme has received two Outstanding Paper Awards and two oral presentations.
Read the full research statement
Developmental intelligence and the science of language modelling
A central theme of my work is how representations and capabilities develop across the stages of a model's training. I am interested in monitoring how representations form, and in evaluating AI systems scientifically rather than only through static benchmarks, as a route to safer systems. Beetle, an open-source small language model suite and learning-dynamics framework built with Paula Buttery, Catherine Arnett and James Michaelov, is my main instrument for this: I use it to study learning dynamics, interpretability and safety under controlled conditions.
Internal representations and generalisation
I study how representations emerge during learning, how their organisation shapes generalisation, composition and transfer, and how we can build better methods for evaluating and monitoring intelligent systems. Recent work asks when independently trained models converge on compatible representations, and what model merging reveals about the geometry of what has been learned.
Human-scale and sample-efficient learning
How can we build more capable systems with substantially less data and computation? I co-organise the BabyLM Workshop and Shared Task at EMNLP 2026 (Budapest) and 2027 (Cancún), and the BabyVLM Workshop at NeurIPS 2026. My master's thesis trained small language models for the BabyLM Challenge using acquisition-inspired curricula on developmentally plausible corpora in several languages. The longer-term aim is to turn a scientific understanding of learning into more sample-efficient AI systems, particularly through better tokenization and a closer study of training data.
Multilingual and comparative cognition
Most people use more than one language, yet computational models of language processing are largely monolingual. I build computational models and small AI systems to study how humans acquire language and other cognitive capabilities, and to compare learning in biological and artificial systems. What can the development of artificial intelligence teach us about cognition, and what can cognitive science contribute to building better AI?
Developmental interpretability and AI safety
If capabilities emerge during training, we should be able to watch them emerge. I am interested in whether developmental signals in a model's internal representations can anticipate behaviour before standard evaluations detect it, and in what that would mean for monitoring, evaluation and the safety of increasingly capable systems.
Collective intelligence and socially situated AI
I want my empirical and theoretical work on small open models to connect with the fast-moving wider AI ecosystem, in an interdisciplinary and stakeholder-centred way. With cognitive scientists and applied machine-learning researchers (for example in language learning), I work on guiding and calibrating LLM behaviour in multi-agent settings, aligning model outputs with user needs and domain-specific tasks, towards AI for decision-making, learning and human–AI collaboration that is better situated socially and culturally.
I am passionate about open-source AI and happy to collaborate and consult with industry partners. In March 2025 our group released PicoLM. The research page sets out each strand with representative papers.
Contributions
Language acquisition · Multilingual models
How does learning history shape linguistic representations? Beetle is a controlled suite of 330 open models (285 bilingual, 45 monolingual) across 13 first languages, five exposure curricula and three data scales. Because the models are matched populations that differ only in one aspect of their linguistic experience, hypotheses about bilingualism, such as the timing of second-language exposure, cross-lingual transfer and attrition, become experiments that are hard to run in larger, less transparent systems.
Research resource · Empirical findings · Open models · project · paper · code & models · EMNLP 2026
More about this contribution
Beetle is built on the PicoLM decoder backbone, so every model shares an architecture and training code and differs only in the data it saw and when. That makes it possible to attribute differences in representation and behaviour to learning history rather than to incidental design choices. The launch essay explains why controlled pretraining is worth the effort.
PicoLM
Studying how capabilities emerge means training many small models and inspecting them checkpoint by checkpoint, infrastructure most researchers rebuild from scratch. Pico is a lightweight, modular framework for hypothesis-driven small-model research: researchers make a targeted change to architecture or training and observe its effect directly, against a suite of standardised baseline models. It now underpins much of my own work and that of a growing group of collaborators.
project · paper · EMNLP 2025 System Demonstrations
Learning from a child-sized budget
Through BabyLM I study which training strategies let a model acquire real linguistic competence from roughly a hundred million words. Earlier work tested whether theories of acquisition can specify finer-grained, language-specific curricula (Less is More); more recent work studies contingent, multi-turn interaction between a teacher and a small "student" model (Teacher Demonstrations) and token-wise gating for vision–language learning (Looking to Learn), both Outstanding Paper Awards. BabyBabelLM extends the paradigm to 45 languages.
Tokenization as a learned allocation of computation
Tokenization is usually treated as fixed preprocessing. I treat it as part of the learning problem: ByteSpan groups predictable bytes into subwords using an external byte-level model, and LangMAP lets the input representation adapt to the language being modelled.
News
- Sep 2026Released Beetle, an open suite of 330 bilingual and monolingual models, and presented it at X-PPL-26 in Bayonne.
- 2026Three main-conference papers at EMNLP 2026, Budapest: Beetle, LangMAP and Cross-Lingual Alignment Without Joint Training.
- 2026Co-organising BabyLM 2026, now with a multilingual track, and the first BabyVLM workshop at NeurIPS 2026.
- May 2026Invited seminars at SheffieldNLP and the University of Copenhagen on sample-efficient and human-scale language modelling.
- Mar 2026The xBLiMPs project, with Catherine Arnett, Paula Buttery and Andrew Caines, funded by the Language Sciences Incubator Fund.
- Jan 2026BabyBabelLM accepted to EACL 2026; two oral papers at OCP23, hosted at Gonville & Caius.
- Nov 2025Two Outstanding Paper Awards at the BabyLM Workshop, EMNLP 2025, Suzhou.
- Mar 2025Released PicoLM, the Cambridge small-language-model and learning-dynamics framework.
Selected research
Reverse chronological. The full publication list and Google Scholar have the complete record.
Beetle: A Bilingual Model Suite for Modelling Second-Language Processing
S Salhan, C Arnett, J Michaelov, P Buttery
EMNLP 2026 | paper | code | project
LangMAP: A Language-Adaptive Approach to Tokenization
C Meister, S Salhan, A Szablewski, P Lesci, P Buttery, T Pimentel
EMNLP 2026 | write-up
Cross-Lingual Alignment Without Joint Training: Do Monolingual Language Models Converge on Universal Representations?
EJ Zhou, S Salhan, C Arnett
EMNLP 2026 | write-up
BabyLM Turns 4 and Goes Multilingual: Call for Papers for the 2026 BabyLM Workshop
L Choshen, R Cotterell, …, S Salhan, …, A Warstadt, EG Wilcox
BabyLM Workshop, EMNLP 2026 | paper
BabyBabelLM: A Multilingual Benchmark of Developmentally Plausible Training Data
J Jumelet, A Fourtassi, A Haga, …, S Salhan, …, A Warstadt, L Choshen
EACL 2026 | project | write-up
Pico: A Modular Framework for Hypothesis-Driven Small Language Model Research
R Diehl Martinez, DD Africa, Y Weiss, S Salhan, R Daniels, P Buttery
EMNLP 2025 System Demonstrations | paper | project
Teacher Demonstrations in a BabyLM's Zone of Proximal Development for Contingent Multi-Turn Interaction
S Salhan, H Gu, D Rooein, D Galvan-Sosa, G Gaudeau, A Caines, Z Yuan, P Buttery
BabyLM Workshop, EMNLP 2025 (Outstanding Paper Award) | paper
Looking to Learn: Token-wise Dynamic Gating for Low-Resource Vision-Language Modelling
BM Ganescu, S Salhan, A Caines, P Buttery
BabyLM Workshop, EMNLP 2025 (Outstanding Paper Award) | paper
What's the Best Sequence Length for BabyLM?
S Salhan, R Diehl Martinez, Z Goriely, P Buttery
BabyLM Workshop, EMNLP 2025 | paper
BLiSS: Evaluating Bilingual Learner Competence in Second Language Small Language Models
Y Gao, S Salhan, A Caines, P Buttery, W Sun
BabyLM Workshop, EMNLP 2025 | paper
Measuring Grammatical Diversity from Small Corpora: Derivational Entropy Rates, Mean Length of Utterances, and Annotation Invariance
F Moscoso del Prado Martín, S Salhan
ACL 2025 | paper
ByteSpan: Information-Driven Subword Tokenisation
Z Goriely, S Salhan, P Lesci, J Cheng, P Buttery
Tokenization Workshop, ICML 2025 | paper
Less is More: Pre-Training Cross-Lingual Small-Scale Language Models with Cognitively-Plausible Curriculum Learning Strategies
S Salhan, R Diehl Martinez, Z Goriely, P Buttery
BabyLM Challenge, CoNLL 2024 | paper
Teaching
- Li18 Computational Linguistics (Linguistics Tripos), lectures on computational syntax. 2025–26.
- L95 Introduction to Natural Language Syntax and Parsing (MPhil ACS / Part III), guest lecturer and teaching assistant. 2024–26.
- MPhil guest lecture, Language Model Evaluation, with Paula Buttery & Fermín Moscoso del Prado Martín. Michaelmas 2024.
- UROP, Small Language Models, with Paula Buttery. Summer 2025.
- Supervisions for the Computer Science Tripos: Machine Learning & Bayesian Inference (II); Artificial Intelligence and Formal Models of Language (IB); Introduction to Probability and Machine Learning & Real-World Data (IA).
Research group
Students I supervise or co-supervise in the Cambridge NLIP group, mostly with Paula Buttery and Andrew Caines.
MPhil students
- Yeji Heo (MPhil ACS 2025–26), cultural knowledge in pretrained language models
- Adam El-Kholy (MPhil MLMI 2025–26), exploration limits of post-training under GRPO
Undergraduate research students
- Ellie Polyakova Reed, Shivan Arora, Mingchuan Zou, Jacy To, Abdullah Akram, Alisa Potomkiva
Former mentees
- Bianca-Mihaela Ganescu (MPhil ACS 2024–25; Outstanding Paper Award, BabyLM 2025)
- Andrzej Szablewski (visiting student; LangMAP, EMNLP 2026) · Ali Kheirkhah (PicoLM intern, 2025)
Collaboration & mentorship
Much of my work is collective. I co-organise the BabyLM challenge and workshop and BabyVLM at NeurIPS, organise the weekly Cambridge NLIP seminar, and develop open research infrastructure through PicoLM. I also teach and supervise across computer science and linguistics (service).
I am interested in collaborating with researchers in machine learning, cognitive science, linguistics, interpretability and AI safety, for example on controlled experiments with small language models, developmental interpretability and model monitoring, multilingual and cross-lingual evaluation, sample-efficient learning, evaluation and calibration, or open-source research infrastructure. I am also happy to talk with industry partners about open models.
Each year I offer MPhil projects in dynamic tokenization, sample-efficient learning, developmental interpretability and bilingual modelling with Beetle, and supervise undergraduate research over the summer. Projects are designed to lead to publishable work. The project areas and 2026–27 proposals (PDF) have details; otherwise, email me.
Writing
About
I'm a PhD candidate in Computer Science at Cambridge and a member of Gonville & Caius College, supervised by Paula Buttery. Before that I studied Computer Science and Linguistics at Cambridge.
I also organise research workshops, build open-source AI systems, and edit Per Capita Media.
Read my extended biography
My work sits between computer science, computational linguistics and cognitive science, and the other fields I draw on supply questions and methods rather than decoration. Language acquisition and developmental psychology suggest which learning trajectories are worth testing and what counts as a plausible data budget. Comparative cognition asks which properties of learning are general and which belong to a particular kind of learner. Computational linguistics and linguistic diversity keep evaluation honest beyond English and beyond high-resource languages. Cognitive science and theories of representation shape what I look for inside models. Philosophy of science informs how I design controlled experiments, and AI safety and governance are why I care whether developmental signals can be monitored at all.
Work from this programme has received two Outstanding Paper Awards at BabyLM (EMNLP 2025) and two oral presentations. Outside research I lead Per Capita Media, an independent Cambridge publication, and have worked in journalism with The One Show, The Sunday Times and BBC Radio 5 Live.
Elsewhere on this site
Publications · CV · Research · Beetle · Teaching & supervision · Talks · Service · Writing · People · About
sas245@cam.ac.uk · Department of Computer Science & Technology, University of Cambridge