Professor James C. R. Whittington and Dr Ruairidh McLennan Battleday, cofounders of stealth AI research laboratory Prefrontal, have released new research addressing one of the major technical challenges facing the next generation of AI: how systems can efficiently process enormous amounts of information.
Their latest paper introduces a method that achieves up to 72.4x faster attention-layer processing at 10m tokens of context — compared with conventional full attention — while maintaining comparable accuracy.
The research follows the publication by DiG-bench, their benchmark examining whether AI systems can genuinely discover unfamiliar rules through experimentation, which reached more than 1m views on social media.
Together, the publications offer an early glimpse into Prefrontal, a British frontier AI laboratory operating in stealth, developing human-inspired AI aimed at accelerating scientific discovery.
Making AI capable of working with millions of tokens
As AI systems move beyond answering questions and towards conducting scientific research, they increasingly need to process vast amounts of information.
A scientific research agent, for example, might need to examine hundreds of papers, follow references between studies, analyse experimental results and connect findings accumulated over many hours.
But processing this information is computationally expensive.
The main bottleneck is attention — the mechanism AI models use to identify relevant information within their context.
As that context grows, the computational cost increases substantially. Prefrontal's new method, Latent Sparse Attention (LSA), addresses this problem by organising information according to meaning rather than its position in a sequence.
Instead of searching through millions of individual pieces of information, the system groups related information into clusters, allowing it to identify relevant material more efficiently.
The results include 72.4x faster attention-layer prefill compared with full attention, 22.7x faster prefill compared with an existing sparse-attention approach and up to 9.2x faster time to first token in an end-to-end model-serving evaluation.
The 72.4-fold improvement applies specifically to attention-layer processing, rather than the performance of an entire AI system.
From a viral discovery benchmark to a frontier AI lab
This new research builds on DiG-bench (Discovery in Games), developed by Battleday, Whittington and an international team of collaborators.
DiG-bench asks a fundamental question: can today's AI systems genuinely discover something new?
Its 70 interactive games challenge AI agents to explore unfamiliar environments, experiment, infer hidden rules and discover how to succeed.
Every game was solved by at least one human participant on their first attempt, while even the strongest AI agents failed to complete a substantial proportion of the challenges.
The findings exposed limitations in the ability of leading AI systems to discover new knowledge through experimentation.
For Prefrontal, DiG-bench and Latent Sparse Attention addresses complementary challenges: enabling machines to discover new information and helping them process the enormous amounts of knowledge required for scientific work.
The researchers behind Prefrontal
Professor James Whittington is a computational neuroscientist and AI researcher at the University of Oxford, where he is a Principal Investigator in the Department of Experimental Psychology.
Having studied physics, medicine and neuroscience at Oxford University, he went on to hold a Sir Henry Wellcome Fellowship at Stanford University.
His research investigates how the brain represents knowledge, learns relationships and generalises across experiences.
Dr Ruairidh Battleday is a British cognitive scientist, AI researcher and entrepreneur. Originally trained in medicine at Oxford, he completed his PhD in Computer Science at Princeton University before conducting research at Harvard.
His research focuses on generalisation, scientific reasoning and the computational principles underlying human learning and discovery.
Building the UK's AI research ecosystem
Beyond Prefrontal, Whittington and Battleday also cofounded Thinking About Thinking, an independent nonprofit research and education organisation established in 2020.
Through scientific conferences, seminars, an international Fellowship and an Ambassador Programme, the organisation connects researchers and emerging talent across institutions including the University of Oxford, the University of Cambridge, UCL, Imperial College London, Stanford, MIT and Princeton.
Later this month, the organisation will host the AE Global Summit 2026: Open Problems for AI, taking place from 25–27 November at Friends House in London.
The three-day event will bring together researchers, AI companies, entrepreneurs, investors and policymakers to explore breakthroughs in AI research, infrastructure, safety, enterprise, entrepreneurship and biomedicine.
The programme includes speakers from Google DeepMind, Microsoft, NVIDIA, Oxford University, Imperial College London, UCL and emerging frontier AI companies, alongside an AI Startup Competition offering a £5,000 prize.
For Whittington and Battleday, the combination of research, a stealth frontier AI lab and an international scientific community reflects a broader ambition: advancing AI while strengthening the research ecosystem needed to translate those advances into scientific discovery.
