Pittsburgh, PA / Founded 2026
Sampling is search.
A diffusion model spends the same budget on every draw, easy or impossible. A reasoning model spends more when the problem is hard, but thinks in one dimension — left to right, one token at a time. Loopflow is building the loop between them: samplers that decide how long to think, revisit their own drafts, and keep only the ones that survive checking.
Or write to martin@loopflowlabs.com
Where we are pointing it
Robot policies that re-plan mid-motion
Action chunks are cheap to sample and expensive to be wrong about. We want a policy that notices the world has diverged from its plan and spends another pass on the part that broke — not the whole trajectory.
Images with structure that has to hold
Schematics, layouts, dense text, physical geometry. Cases where the picture is checkable, and a sampler that can check its own draft beats one that draws faster.
Discrete problems, denoised in parallel
Scheduling, layout, program synthesis. Whole candidate structures refined at once, with iterations allocated where the constraint violations actually are.