Probabilistic Tiny Recursive Model
Summary
Probabilistic Tiny Recursive Model (PTRM) is the training-free inference extension to Tiny Recursive Model introduced by Probabilistic Tiny Recursive Model. It injects Gaussian noise at every deep-recursion step, runs multiple latent reasoning trajectories in parallel, and uses the pretrained TRM Q head to select one answer.
Role In The Wiki
Use this page as the object card for PTRM. The source page carries the detailed evidence, social-claim audit, cost calibration, limitations, and agenda mapping.
PTRM is especially useful for separating:
- proposal coverage: whether any stochastic rollout reaches a correct candidate, measured by pass@;
- selection quality: whether the internal Q head identifies that candidate, measured by best-Q@;
- parallel width: how many trajectories are sampled;
- sequential depth: how many recursive refinement steps each trajectory runs.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in probabilistic-tiny-recursive-model-2026 rather than duplicating verdict rows here. At the entity level, PTRM is adjacent machinery for parallel inference-time compute and latent-state exploration. It does not yet provide calibrated multiple plausible futures, numeric time-series state tracking, or action-conditioned world-model evidence.
Evidence
Official Artifacts
- Preprint: arXiv 2605.19943v1
- Official project page: Probabilistic Tiny Recursive Model
- Code/checkpoint status: the official project page says code is coming soon; no official PTRM implementation or checkpoint was verified as of 2026-07-17.