Research Interests

We are challenge the scale-is-all-you-need paradigm by developing Small Language Models (SLMs) that are more data-efficient, easier to explain and cognitively plausible.

Buttery Group's Research Themes

• Sovereign AI: Optimization, design, and data-efficiency of Small Language Models (SLMs) for local, secure deployment
• Explainable AI: Distinguishing between intrinsic model interpretability and stakeholder-focused explanations

For Buttery Group publications see acl anthology and scholar.

1. Sovereign AI with Small Language Models

We develop architectures that allow stakeholders to own their intelligence rather. Our work focuses on:

• The Pico Framework: We have developed a modular environment for training high-performance Small Langauge Models on affordable hardware.
• Data Sovereignty: We aim to avoid the risks data leakage and monoculture by building performant bespoke models from currated data sets
• Model Evaluations: We investigate stability, convergence and learning dynamics to guide design of models that are performant without massive scale. We critique and rethink evaluation paradigms so that evaluation is appropriate for models and tasks

2. Explainable AI

We maintain a technical distinction between model properties and user-facing outputs:

• Interpretability: Interpretability is an intrinsic property of a model. We investigate mechanistic interpretability with a view to understanding (and redesigning for) bottlenecks.
• Explanations: We treat this as a translation task. We build frameworks and mechanisms to convert interpretable model states into appropriate formats for specific stakeholders (e.g., auditors, regulators, or end-users).

Cognitive Inspiration

Many of machine learning approaches are grounded in computational Psycholinguistics. By understanding how humans learn language efficiently from noisy data, we design curricula and architectures that mimic these biological efficiencies.

Key Applications

• FinTech and RegTech: Applying specialized SLMs to financial and regulatory text to automate compliance monitoring and policy verification without data leakage (in collaboration with initiatives like RegGenome).
• Modelling for Code: Developing small, specialized models optimized for code generation, and formal protocol verification.
• Automated Teaching and Assessment : Optimising learning journeys for students, teachers and examiners (the Buttery Group leads the Automated Language Teaching and Assessment (ALTA) Institute).
• Low-Resource Languages: NLP tools for endangered languages and non-canonical text (e.g., aphasic speech, learner language).