Postdoctoral Research Associate in Adaptive and Efficient LLM Architectures (ERC project AToM)

The AToM project (Adaptive Tokenization and Memory in Foundation Models), an ERC Starting Grant led by Edoardo M. Ponti at Imperial College London, is recruiting a Research Associate for 2 years, starting around winter 2026/27 (flexible).

We aim to build models that learn to compress what they perceive and remember, to achieve lifelong memory, extended horizons, and >10x speedups. My lab has prototyped latent tokenization and adaptive-size memory compression, and released retrofitted open-weight models with NVIDIA (Qwen3-8B-DMS-8x) and Ai2 (Bolmo-7B). We aim to build on these foundations to scale them up and explore new research directions:

1. Latent tokenization of raw, multimodal data
2. Permanent adaptive-size memory and hierarchical sparse attention
3. Retrofitting state-of-the-art models, with open recipes, data, and checkpoints
4. Applications in inference-time hyper-scaling (for reasoning and agentic workflows) and long-horizon world modelling

What we offer:

• Compute: project-owned B200 GPUs and cloud credits + departmental H200s
• Close collaboration with frontier labs
• Funding for two conferences a year
• Freedom to set an independent research agenda and co-supervise PhD/MSc students.

Some more info:

• Application deadline: 30 September 2026
• Starting date flexible around winter 26/27
• Salary: GBP 50,733 to 59,484 per annum (Research Assistant scale, GBP 45,399 to 48,876, if the PhD is not yet awarded).

Application process:

• Please apply through the listed job posting, attaching a full CV with a publication list and a research statement of a maximum of 2 pages (please focus it on project-related research).
• Candidates are also welcome to fill in the expression-of-interest form: https://t.co/Tn2cosHB06
• Shortlisted candidates will be invited to an interview in October, with a research talk and a technical discussion.