Research Associate in Adaptive and Efficient LLM Architectures

Overview

In this postdoc, you will lead research on efficient and adaptive architectures for AI models within the ERC project AToM. You’ll develop and release open-weight LLMs with adaptive memory, latent tokenization, and other efficiency techniques, collaborating with Dr. Edoardo Ponti. The role focuses on advancing long-context understanding, multimodal world modelling, and scalable inference while contributing to a world-leading research agenda. This is a fully funded, impact-oriented position at Imperial College London.

Pay / Benefits

• conference travel (2 international conferences per year)
• extensive compute resources (AToM GPUs, A100s, H200s)
• career support for researchers
• sector-leading salary and remuneration
• 43 days off per year
• generous pension schemes

Responsibilities

• Design and develop novel AI architectures for efficient LLMs
• Retrofitting, post-training, and evaluating state-of-the-art open-weight models
• Conduct independent and collaborative research within the group
• Publish results in top-tier conferences and journals (e.g., NeurIPS, ICML, ICLR, ACL, EMNLP, Nature)
• Implement research ideas using PyTorch/JAX, HuggingFace transformers/diffusers, and efficient kernels (Triton/CUDA)
• Contribute to projects on adaptive memory, latent tokenization, sparse attention, long-context reasoning, and multimodal world modelling
• Mentor PhD and MSc students; participate in group meetings and maintain project resources
• Prepare grant proposals and participate in collaborative research initiatives
• Deliver tutorials at conferences and summer schools and present at international venues

Key requirements

• Strong track record in AI/ML/NLP conferences and journals
• Excellent coding skills; strong foundations in information theory, linear algebra, calculus
• Experience with LLM training, evaluation, RLVR, PEFT, quantisation, tensor/data parallelism
• Familiarity with CUDA kernels and/or Triton; experience with inference engines (vLLM, SGLang)
• Proficiency in PyTorch/JAX
• Fluent English communication skills; ability to contribute to group culture
• PhD in computer science or equivalent experience

• Self-driven and motivated
• Strong collaboration and communication
• Ability to present and mentor others
• LLM training and evaluation
• RLVR (reinforcement learning for VL tasks)
• PEFT (parameter-efficient fine-tuning)

…