Flow Reasoning Models turn flow matching into iterative reasoners, hitting 99.5% on Sudoku-Extreme
alec_helbling · x · 2026-10-26
A new arXiv paper introduces Flow Reasoning Models (FRMs), extending flow matching to structured reasoning. FRMs self-condition on their own past outputs, turning one-shot denoising into iterative solution refinement so they can make and revise interdependent decisions in parallel—unlike autoregressive models that can't revise earlier choices. To fix exposure bias at greater recurrent depth, the authors propose Fixed-Point Forcing (FPF), training on states produced by the model's own inference dynamics while preserving the flow-matching objective. FRMs reach 99.5% on Sudoku-Extreme, 100% on Zebra, and 99.9% on Maze-Unique, with higher peak accuracy than evaluated masked-diffusion baselines on Sudoku-Extreme.
More from Research
- Laying out DeepLIFT as a standalone computation graph: from sequence to attributions directly — jmschreiber91 · 2026-10-26
- Speculative Diffusion Paper Maps the Speed-Quality Pareto Frontier via Two Tradeoffs — UCLA · 2026-10-26
- Git Records Fall Short for Auditing AI's Role in Research Repositories, Paper Finds — Koutian Wu · 2026-10-26
- Fused DeepLIFT explained: why attribution hooks are cheap for Borzoi/AlphaGenome but hurt small models — jmschreiber91 · 2026-10-26
- New paper: AI agents negotiate badly under uncertainty and break deals on purpose — menhguin · 2026-10-26
- Cherimoya v0.3.1 Ships Fused DeepLIFT, ~7x Faster Attribution for Genomics Models — jmschreiber91 · 2026-10-26