Our research
Our research themes
Our products integrate generative AI, and part of this research works one level down, on the mechanisms of generation and reasoning themselves, while the other part addresses the detection, acceleration and decision problems that run through civil systems and defence alike.
The price of knowing
An inspection immobilises the asset, a high-fidelity simulation burns days of compute, and a destructive test consumes the part.
Read the brief Machine-native reasoningAbstract CoT, the price of thinking aloud
Reasoning written in human sentences is billed by the token, and its length is unpredictable.
Read the brief Certifiable embedded AIEmbedded AI acceleration
From a trained model to FPGA hardware that can be certified for safety-critical, regulated systems.
Read the brief Trust & media integrityDetecting AI-generated video
Finding the traces generation leaves behind, and staying robust as the generators keep improving.
Read the briefResearch notes
What we have published
Decision environments where information has a price
Where paying for a measurement wins, and where free data is already enough.
Can a model reason in a language of its own?
One test decides whether a model's internal code carries its reasoning or merely decorates it.
A reasoning channel must break when you scramble it
A 26.5M-parameter model whose internal code was certified as the sole causal path, with the run records behind it.
Three ways a reasoning channel fails silently
The failure mode that passes every dashboard while carrying none of the reasoning, and the counterfactual test that catches it.
Does planning help a model design a molecule?
What a continuous bottleneck costs in controllability against a discrete one, measured at roughly a factor of six, and a plan that never beat direct conditioning.
Discuss a research problem with us
If one of these problems is yours, or close to it, we are open to focused conversations with partners, integrators and public bodies.
Write to us