The University of Edinburgh, through the ERC-funded project "Adaptive Tokenization and Memory in Foundation Models for Efficient and Long-Horizon AI" (AToM), is offering a fully funded 3.5-year PhD studentship and a 2-year postdoctoral position starting autumn 2026. The project, led by Assistant Professor Edoardo Ponti, focuses on advancing foundation models in artificial intelligence by developing adaptive tokenization and memory mechanisms. This research aims to make AI models more efficient, reducing energy consumption and carbon emissions, while enabling longer-term memory and improved reasoning capabilities.

Successful candidates will join Edinburgh NLP, a world-leading research group within the School of Informatics, and work on cutting-edge topics such as end-to-end tokenization, learned memory compression, and inference-time hyper-scaling. The project builds on recent advances, including dynamic memory compression and key–value cache compression for large language models, as demonstrated in recent publications and open-source releases. The research environment is highly interdisciplinary, with opportunities to collaborate with leading experts and access state-of-the-art computing resources and travel funding.

Applicants for the PhD position should have a strong academic background in informatics, computer science, artificial intelligence, or related fields, with experience in programming. Degrees in engineering, linguistics, mathematics, philosophy, physics, or psychology may also be considered. English language proficiency is required. The postdoctoral position requires a relevant PhD and research experience in related areas. Both positions are fully funded, covering tuition, stipend, and research expenses.

To apply for the PhD, candidates should submit an application via the ILCC programme at the University of Edinburgh, selecting Edoardo Ponti as the proposed supervisor. For the postdoc, interested applicants should contact Edoardo Ponti directly by email. The application deadline for School of Informatics funding is 18 December 2025. For more information, see the ILCC programme page and related project links.

Keywords: Natural Language Processing, Foundation Models, Adaptive Tokenization, Memory Compression, Efficient AI, Long-Horizon AI, Transformer Models, Deep Learning, Artificial Intelligence, Computational Linguistics.