FRM
Summary
FRM, short for Flow Reasoning Model, is the discrete-flow structured-reasoning framework introduced by Flow Reasoning Models. It combines self-conditioned iterative denoising, perturb-and-resolve stability scoring, verify-and-restart test-time scaling, and localized FlowDPO training on self-mined wrong completions.
Role In The Wiki
Use this page as the object card for the model and method family. The source page carries the detailed evidence, claim audit, limitations, and agenda mapping.
FRM is especially useful for distinguishing three fixed-point interfaces that are often conflated:
- convergence as a halting signal;
- stability under perturbation as a candidate-ranking signal;
- preference training that reshapes which fixed points the sampler reaches.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in frm-2026 rather than duplicating verdict rows here. At the entity level, FRM is adjacent machinery for dynamic compute, candidate-state verification, and self-corrective generation. Current evidence is limited to discrete checkable puzzles and does not establish calibrated multivariate time-series state prediction, multiple plausible futures, or action-conditioned world modeling.
Evidence
Official Artifacts
- Preprint: arXiv 2606.29150v1
- Code/checkpoint status: no official repository or released checkpoint was verified as of 2026-07-16.