Flow Reasoning Models: recurrent flow architecture hits 99.5% on Sudoku-Extreme with 44x fewer FLOPs
alec_helbling · x · 2026-09-03
A new arXiv paper introduces Flow Reasoning Models (FRMs), a framework for structured reasoning that applies continuous flows to discrete structured outputs with recurrent self-conditioning refinement.
- How it works: self-conditioning a flow model on its own past outputs turns one-shot denoising into iterative solution refinement, letting the model make and revise interdependent decisions in parallel.
- Key fix: conventional self-conditioning breaks down at greater recurrent depth due to exposure bias; Fixed-Point Forcing (FPF) trains on states produced by the model's own inference dynamics while preserving the flow-matching objective. Multiple recurrent passes lift Sudoku-Extreme from 30% to 99%.
- Results: 99.5% / 100.0% / 99.9% solve rates on Sudoku-Extreme, Zebra, and Maze-Unique; on Sudoku-Extreme FRMs match EqR's 98.7% with 44x fewer FLOPs and reach a higher 99.5% peak accuracy, outperforming HRM, TRM, and masked diffusion on inference efficiency.
- Perspective: correct answers are fixed-point attractors of the self-conditioned flow's dynamics.
Related event: Flow Reasoning Models Solve Sudoku With 44x Less Compute(4 posts)→
More from Research
- Action Chunking Boosts Contrastive RL Even in Fully Online RL, Study Finds — ben_eysenbach · 2026-09-03
- Coding models are running out of data — PL researchers propose 'intent computing' as the fix — LingmingZhang · 2026-09-03
- davidad Backs Call to Ban Naive RLVR: 'Everything Should Be Model-Graded' — davidad · 2026-09-03
- Computerphile Deep Dive: How Watermarks Track AI-Generated Content — Computerphile · 2026-09-03
- TrafficLab 3D builds digital-twin traffic visualizations from CCTV footage and Google Maps — tom_doerr · 2026-09-03
- SOCO benchmark debuts at ECCV 2026: 1M+ pairs probe how vision models grasp object structure — HirokatuKataoka · 2026-09-03