Flow Reasoning Models Turn Flow Models Into Recurrent Reasoners, Hitting 99.5% on Sudoku-Extreme

alec_helbling · x · 2026-09-17

Researchers introduce Flow Reasoning Models (FRMs), a structured-reasoning framework that applies continuous flows to discrete structured outputs and recurrently refines predictions via self-conditioning. Unlike autoregressive models that commit sequentially or masked diffusion models needing careful decoding, FRMs make and revise interdependent decisions in parallel, scaling test-time compute by looping longer—each pass gets a local loss, no backprop through time. A new Fixed-Point Forcing (FPF) scheme trains on states from the model's own inference dynamics to fix exposure bias at deeper recurrence. FRMs hit 99.5%, 100.0% and 99.9% solve rates on Sudoku-Extreme, Zebra and Maze-Unique, with higher peak accuracy than masked-diffusion baselines on Sudoku-Extreme.

Related event: Flow Reasoning Models Turn Flow Models into Recurrent Reasoners(3 posts)→

Original post →

More from Research

Research channel →