Edoardo Ponti

Imperial College London is recruiting PhD students and postdocs for research with Edoardo Ponti in the Department of Computing. The project focuses on foundation models , adaptive tokenisation , memory compression , and efficient large language models, including KV cache compression , long-context modelling, and energy-efficient AI. The advertised research direction is part of an ERC Starting Grant project called AToM FMs (adaptive tokenisation and memory). The lab aims to study whether foundation models can learn end-to-end to adapt the length of their representations rather than relying on fixed tokenisation and growing memory. The post highlights prior work on self-compressing Transformer architectures, dynamic memory compression, latent autoregressive tokenisation, and long-horizon modelling. Position types: PhD and postdoc openings are explicitly mentioned. The positions are stated to be fully funded , and starting dates are flexible. Eligibility and requirements: For the PhD route, Imperial states that applicants are normally expected to hold a First Class or Distinction Masters degree, or equivalent, in a relevant scientific or technical discipline such as computer science or mathematics. Applicants with only a bachelor’s degree are usually not considered directly for the PhD. The post does not list formal postdoc requirements, but strong alignment with NLP, machine learning, and efficient transformer research is implied. How to apply: Interested candidates should email or DM Edoardo Ponti. For the PhD application, Imperial directs applicants to the Computing Research degree and the online application system. The department notes that there is currently no application fee and that overseas candidates needing funding should apply by the December funding deadline. Deadline information: For 2026 entry, Imperial says applications open in October 2025 and can be made throughout the year. The funding-related deadline highlighted in the department page is 15 April 2026 for one of the listed rounds, with earlier deadlines also shown for other rounds. The post itself does not give a single fixed deadline, so applicants should check the department page and apply early. Research keywords: foundation models, adaptive tokenisation, memory compression, efficient transformers, NLP, long-context learning, continual learning, world modelling, inference-time scaling.

Imperial College London is recruiting for research positions in adaptive tokenization and memory in foundation models as part of Edoardo Ponti’s ERC Starting Grant project, AToM (Adaptive Tokenization and Memory in Foundation Models). The call includes 1 postdoc / Research Associate and 2 fully funded PhD students . The research agenda focuses on latent tokenization , permanent adaptive-size memory , hierarchical sparse attention , retrofitting state-of-the-art models, and applications in inference-time hyper-scaling , reasoning, agentic workflows, and long-horizon world modelling. The lab highlights prior work such as Dynamic Token Pooling , Bolmo , and Dynamic Memory Compression , with open-weight model releases in collaboration with NVIDIA and Ai2. Successful candidates will work with project-owned B200 GPUs, departmental H200s, cloud credits, and receive funding for two conferences per year. The postdoc is a 2-year Research Associate role with flexibility to set an independent research agenda and co-supervise PhD/MSc students. The PhD positions are approximately 3.5 years and are applied for through Imperial Computing admissions on a rolling basis. Deadline: 30 September 2026 for the postdoc. The PhD applications are rolling. How to apply: Use the expression-of-interest link for the postdoc and the Imperial Computing admissions route for the PhD positions. Review the details link for full instructions.

Imperial College London is recruiting for 1 postdoc and 2 fully funded PhD positions in the AToM ERC research project led by Edoardo Ponti in the Imperial College Computing community. The project focuses on artificial intelligence and machine learning research around models that learn to compress what they perceive and remember , enabling permanent memory , long-horizon reasoning , and potential 10x speedups . Related topics mentioned include latent tokenization in decoders, learnable adaptive-size memory , inference-time hyper-scaling , and long-horizon world modelling . Funding and resources are strong: the positions are described as fully funded , with access to project-owned B200s , cloud credits , H200s from Imperial Computing, and funding for 2 conferences per year . Eligibility notes from the post: both overseas and home students are welcome . Starting dates are flexible. The postdoc deadline is 30 September 2026 , while PhD applications are accepted on a rolling basis . To apply, first complete the interest form and then submit the formal application through the links provided inside the form. The project website is linked in the post for more details.