Visual-reasoning models can solve problems by generating the answer image directly, and flow matching—an iterative sampling method—seems like a natural fit: each sampling step could act as one reasoning step. The authors of this paper show that this intuition fails for the standard 'vanilla' flow-matching sampler. Under mild regularity conditions, they prove that the sampler's entire trajectory can be simulated by a constant number of rounds of parallel model evaluations, so adding sampling steps cannot increase what a constant-depth transformer backbone can express.

The bottleneck disappears when the model is given self-conditioning: it receives its own previous estimate of the final image as an extra input. The authors prove that with this feedback, a constant-depth backbone gains expressive power equivalent to space-bounded Turing machines. In other words, the extra information carried between steps is what allows more steps to mean more computation.

Controlled experiments on several visual-reasoning tasks support the theory. Flow matching with self-conditioning achieved higher accuracy and greater reasoning depth than vanilla flow matching, and its performance continued to improve as the number of sampling steps increased. The takeaway is a simple design principle for future work: carry more information between sampling steps, so that additional steps translate into deeper reasoning.