Selected work on structural sequence modelling, self-improving systems, and world models.
A complete publication list is available on Google Scholar .
Preprint
NoteBank expands the long-context capability of frozen linear attention models with sparse episodic memory and native replay, without additional training.
† Equal contribution
arXiv preprint
A system-level survey of self-improving agents, unifying updates to foundation models and to prompts, memory, tools, and executable control logic.
ICLR 2026Oral
HGM searches over self-modifying coding agents using the performance of their descendants to estimate future improvement potential, rather than selecting only for current ability.
ICLR 2026
SlotSPE turns histology and genomic profiles into distinct prognostic-event slots, then models their interactions for interpretable multimodal survival prediction.
† Equal contribution
ICLR 2025
FACTS combines graph-structured recurrent memory, learned routing, and selective state-space propagation to model interacting factors across spatial-temporal sequences.
PhD thesis · University of Edinburgh
Unsupervised generative methods for recovering objects and scene structure: multi-view learning with MulMON, latent duplicate suppression, and spatial-temporal factorization with DyMON.
NeurIPS 2021
DyMON learns object-centric representations of dynamic scenes without supervision, separating camera motion from object dynamics to support queries across space and time.
NeurIPS 2020Spotlight
MulMON accumulates evidence across multiple views into consistent object representations, resolving single-view ambiguities and predicting images and object segmentations at new viewpoints.
