Flow Reasoning Models hit ~100% solve rates on hard structured reasoning benchmarks
burny_tech · x · 2026-09-12
A thread collecting concurrent work on structured reasoning: Flow Reasoning Models (FRMs) and Diffusion as a Training Curriculum for Timestep-Free Iterative Reasoning.
FRM (arXiv:2606.29150) key ideas:
- Existing architectures struggle: autoregressive models commit sequentially and can't revise earlier decisions; masked diffusion models need careful decoding schemes for interdependent predictions.
- FRMs adapt continuous flows over discrete structured outputs, using self-conditioning on past outputs to turn one-shot denoising into iterative solution refinement — making and revising decisions in parallel.
- To fix exposure bias at greater recurrent depth, they propose Fixed-Point Forcing (FPF), training on states produced by the model's own inference dynamics while keeping the standard flow-matching objective.
- Results: 99.5%, 100.0%, and 99.9% solve rates on Sudoku-Extreme, Zebra, and Maze-Unique, with higher peak accuracy on Sudoku-Extreme than evaluated masked-diffusion and sparse methods.
The author sees substantial utility in these approaches and is excited about their development.
More from Research
- 4 sink tokens + 64-token window matches distilled linear attention, no training needed — burny_tech · 2026-09-14
- Oak Ridge NL builds fully automated AI system that assembles molecules atom-by-atom for 25+ hours — jiqizhixin · 2026-09-14
- CMU launches Benchmark Radar, a living search engine for AI benchmarks — CarnegieMellonU · 2026-09-14
- Two weeks, two courses: writing a flow matching post by timing what to hide — ariG23498 · 2026-09-14
- Carl Feynman explains why Navier-Stokes became a Millennium Prize problem — burny_tech · 2026-09-14
- UMass AutoIndex Turns Chunking Into Programs an LLM Writes, Not Configs You Tune — mrdrozdov · 2026-09-14