Yaseen M. Osman

MINDS CDT doctoral researcher | University of Southampton | MINDS CDT, 2023

Making large language models more efficient, private and accessible

Yaseen M. Osman is investigating how large language models can run more efficiently on lower-compute devices, with potential benefits for privacy, sustainability and wider access to AI.

Profiles: LinkedIn | ORCID

Research at a glance

Research focus	Efficient large language models
MINDS theme	Embedded Artificial Intelligence
Technologies	Transformer-based language models, edge and on-device AI
Methods	Activation analysis, quantisation, pruning, hardware and kernel optimisation, active learning
Potential wider relevance	More private, lower-compute and more sustainable AI
Expected completion	2027

Running large language models with less computing power

Large language models, the technology behind many modern AI assistants, need substantial computing power. As these models grow, so do the cost, energy and server infrastructure needed to run them. Most are run on remote servers, which means data must leave the device it was created on, raising questions about privacy and about who can access the technology.

Yaseen studies activation patterns: the internal signals produced inside a transformer-based language model as it processes text. His research investigates whether these patterns can reveal how models work in ways that help them run on lower-compute hardware, such as personal devices, while keeping the performance that makes them useful.

Why efficient language models matter

Research approach

The research combines analysis of what happens inside language models with established techniques for reducing their computing needs. These include quantisation, which stores a model’s numbers at lower precision; pruning, which removes parts of a model that contribute little; hardware and kernel optimisation, which tunes how calculations run on a given chip; and active learning, which aims to choose the most useful data or examples. The effect of each technique depends on the model and the hardware, so each is tested rather than assumed to help.

Research keywords: Large language models, efficient AI, edge AI, transformer models, activation patterns, quantisation, pruning, active learning

Research and outputs

Featured output: arXiv preprint, 2026

Activation-Based Active Learning for In-Context Learning: Challenges and Insights, Yaseen M. Osman, Geoff V. Merrett and Stuart E. Middleton. First submitted to arXiv in June 2026 and revised in September 2026.

Language models can learn a new task from a few examples included in a prompt. This study tested whether activation signals inside two 3-billion-parameter models could help select which examples to include. The signals did not reliably predict which examples improved performance, a finding that helps direct future work towards new ways of interpreting what happens inside these models.

Read the preprint on arXiv

Research profiles

• Yaseen M. Osman on ORCID
• Yaseen M. Osman on LinkedIn

Policy contribution

Yaseen contributed written evidence to a UK Parliament committee inquiry. This is a policy submission, separate from his peer-reviewed research outputs.

Read the written evidence on the UK Parliament website (PDF)

Outreach and public engagement

Talking River (Talking Itchen)

Talking Itchen connects water-quality sensors in the River Itchen to a language model, so that visitors can ask questions about the health of the chalk stream. Yaseen developed the project with SustAI postgraduate researchers Angelina Spilnyk and Andrei Șontea. It began as the runner-up idea at the 2025 MINDS and SustAI Innovation Camp.

Yaseen worked on the language model and taught the AI sessions. Over eight weekly sessions, he took 19 children from the Young Guardians of the River Itchen from a simple one-dimensional language model up to the scale used in real systems. Seven of the children went on to present their work at the Southampton Science and Engineering Festival. Yaseen presented Talking Itchen at the British Science Festival on 19 September 2026.

Talking Robot (Rosie and Reachy)

Talking Robot is a conversational robot project Yaseen worked on with MINDS CDT Enterprise Fellow Tommy Tham. He helped run its stand at the British Science Festival on 19 September 2026.

Related stories

• Yaseen Osman on teaching children how language models work, and giving a river a voice
• Robots, rivers and neurons: our AI researchers meet the public at the British Science Festival