FPRM

Summary

FPRM, short for Fixed-Point Reasoning Model, is the looped Transformer reasoning model introduced by Fixed-Point Reasoners. It uses fixed-point convergence of the hidden state as its halting mechanism and combines pre-norm layers with learned residual scaling to keep deep looped computation trainable.

Role In The Wiki

Use this page as the object card for the model/method. The source page carries evidence details, limitations, and the time-series/world-model transfer boundary.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in fprm-2026 rather than duplicating verdict rows here. At the entity level, FPRM is most relevant as an adjacent dynamic-compute and latent-state convergence mechanism: it suggests that a recurrent-depth model can use its own state update residual as a stopping rule, but current evidence is symbolic reasoning rather than multivariate time-series or action-conditioned world-model evidence.

Evidence

• Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers

Official Artifacts

• Preprint: arXiv 2606.18206
• Workshop version: ICML 2026 AdaptFM OpenReview poster
• Official code: nilskiKonjIzDunava/fprm
• Official checkpoints: fixed-point-reasoners/fprm

Related Pages

• Hierarchical Reasoning Model
• Tiny Recursive Model
• GRAM
• LoopFormer
• Parcae
• Looped Transformers And Test-Time Memory
• Efficient Recurrent Sequence Models
• Time-Series Scaling And Efficiency
• Foundation Time-Series Model Research Agenda