Logarithms Labs
Reinforcement learning for critical domains.
“The fox knows many things, but the hedgehog knows one big thing.”
Archilochus
Frontier models are remarkable generalists. They can write, code, search and reason across fields. But broad ability becomes uneven when the work turns specialised and the cost of a plausible mistake becomes real.
Critical domains demand more than a fluent answer. A model must understand the task, use the right tools, recognise uncertainty and produce work that experts can inspect and defend.
That capability is trained. It comes from realistic tasks, rewards that preserve professional judgement, environments where actions have consequences, and human data that records how experts decide.
Logarithms Labs builds those training systems. We work at the uneven edge of model capability: the domains where models are almost useful, but not yet reliable enough to trust.
Our research begins with reinforcement learning for critical work, including spatial intelligence and complex professional workflows. We turn hard-to-measure capability into repeatable experience, rigorous evaluation and better model behaviour.
The training system
Teach the work, not the appearance of it.
A model learns inside a controlled world: it receives a task, acts through tools, changes state, receives expert-aligned reward and is tested against cases it has not seen.
Research