Dayyán O’Brien

Email: Dayyan.OBrien@ed.ac.uk

Website: https://dayyanobrien.github.io/

LinkedIn: https://www.linkedin.com/in/dayyano/

Research keywords: LLM reasoning, compositionality, interpretability

Bio:

Dayyán, originally from Ireland, moved to Edinburgh in 2019 to pursue an Integrated Masters in Informatics. For his honours work, he completed two Honour’s Projects on numeric commonsense across languages, both classified as outstanding, under the supervision of Mirella Lapata. He then spent over a year with the StatMT group at the University of Edinburgh, contributing to five papers, two as first author. His research focused on multilingual reasoning, including the release of a parallel document-level dataset spanning 50 mid-resource languages and a mathematical reasoning task designed to test inductive reasoning in language models without data leakage.

PhD research:

Language models generate fluent text but can struggle with compositionality, the ability to build complex ideas from simpler parts. This capacity is central to human language and reasoning, yet remains poorly understood in LLMs. Dayyán’s research examines compositional structure in conversation, studying how models assemble and integrate ideas across turns. By analysing the representations and processes involved, he aims to identify principles that explain and improve compositional reasoning in LLMs.

Supervisors: Emily Allaway