Postdoc & PhD Opportunities: Efficient & Adaptive AI Architectures at Imperial College London

Want to help build a path toward greener, more capable generative AI — one that's already resulted in real open-weight model releases with NVIDIA and AI2? The AToM project at Imperial is recruiting 1 postdoc and 2 PhD students.

If you work on LLM efficiency, memory, tokenization, or inference-time scaling, this ERC-funded project offers serious compute, strong industry partnerships, and a track record of shipping retrofitted frontier models — not just papers.

About the Project

AToM (Adaptive Tokenization and Memory in Foundation Models) targets a fundamental inefficiency in dominant foundation model architectures like Transformers, where the length of internal representations bottlenecks both prompt processing and output generation. Led by Edoardo Ponti at Imperial College London's Department of Computing, the project has prototyped a new class of architectures that learn to compress sequences of internal representations end-to-end, redefining the model's "atomic units" for processing and memorizing information — yielding 8x speedups without accuracy degradation, and unlocking longer effective context, lifelong memory, and inference-time hyper-scaling for reasoning-heavy tasks.

In its first year, the project has already released two retrofitted open models: Qwen3-8B-DMS-8x with NVIDIA (adaptive memory, 8x KV cache compression) and Bolmo-7B with AI2 (latent tokenization).

Research Directions

Positions will focus on one or more of the project's four foci:

1. Latent tokenization — end-to-end learned segmentation of raw data into hierarchical units

2. Permanent adaptive-size memory and sparse attention — learned memory compression for lifelong learning and long-context inference

3. Retrofitting SOTA models into adaptive architectures, with public release of data, recipes, and models

4. Inference-time hyper-scaling and long-horizon world modelling — freeing up compute for reasoning-intensive tasks and multimodal planning/simulation

What's on Offer

Extensive compute, including the project's own B200 GPUs and cloud credits, plus H200s from the Department of Computing and Faculty of Engineering; close collaboration with industry and academic partners like NVIDIA and AI2; and funding for two international conferences per year. The postdoc role additionally offers freedom to define an independent research agenda within the project, co-supervise PhD and MSc students, and deliver tutorials at conferences and summer schools.

Postdoctoral Position

• Title: Research Associate in Adaptive and Efficient LLM Architectures

• Contract: Full-time, fixed-term for 2 years, starting around winter 2026/27 (flexible)

• Requirements: A PhD in computer science, a strong publication record in ML/NLP venues, and hands-on experience with LLM training and evaluation in PyTorch or JAX; familiarity with CUDA or Triton kernels and inference engines is desirable.

• Deadline: September 30, 2026

👉 Official Imperial job posting

PhD Studentships (2 positions)

• Duration: 3.5 years, fully funded (covering overseas PhD fees)

• Start date: Flexible

• Process: Reviewed on a rolling basis — applicants are contacted before the formal deadline if promising

How to Apply

Both roles start with the same expression-of-interest form:

👉 Expression of Interest form (for both PhD and postdoc candidates)

For the postdoc, also apply formally through the official Imperial posting with a full CV and a research statement. PhD applicants should follow Imperial's Department of Computing PhD application guidelines in parallel, naming Edoardo Ponti as proposed supervisor.

Meet the PI

Edoardo M. Ponti
Department of Computing, Imperial College London
🔗 AToM project page

For more personalized tips or templates contact our team.

www.applywithmentors.com