Professor James C. R. Whittington and Dr Ruairidh McLennan Battleday, cofounders of the British AI research lab Prefrontal, have published a method designed to help AI systems process very large amounts of information more efficiently. The lab says its Latent Sparse Attention (LSA) approach delivers up to 72.4 times faster attention-layer prefill than conventional full attention at a context length of 10 million tokens, with comparable accuracy.
Attention is the mechanism AI models use to identify relevant information in their context. Prefrontal’s researchers say its computational cost rises substantially as that context grows. LSA groups information by meaning, rather than searching through individual items in sequence, so the system can locate relevant material within clusters.
The paper also reports prefill speeds 22.7 times faster than an existing sparse-attention approach and up to 9.2 times faster time to first token in an end-to-end model-serving evaluation. The 72.4-fold figure applies to attention-layer processing, not the performance of an entire AI system.
The work follows DiG-bench, a benchmark developed by Battleday, Whittington and international collaborators to test whether AI agents can discover unfamiliar rules through experimentation. Its 70 interactive games require agents to explore environments, run experiments and infer how to succeed. Each game was solved by at least one human participant on a first attempt, while the strongest AI agents did not complete a substantial proportion of the challenges.
Prefrontal says the projects address related problems: whether machines can discover new information and how they can process the large body of knowledge needed for scientific work. The lab is developing human-inspired AI with the aim of accelerating scientific discovery.
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. Battleday is a cognitive scientist and AI researcher who studied medicine at Oxford, completed a computer science PhD at Princeton University and later conducted research at Harvard.
The two researchers also cofounded Thinking About Thinking, an independent nonprofit research and education organisation established in 2020. It connects researchers and emerging talent through events and programmes. The organisation is scheduled to hold the AE Global Summit 2026: Open Problems for AI in London from 25 to 27 November.
Sources: Sifted.