GRAM

Summary

GRAM, short for Generative Recursive reAsoning Models, is the probabilistic recursive-reasoning framework introduced by Generative Recursive Reasoning. It turns deterministic recursive latent-state refinement into stochastic multi-trajectory computation.

Role In The Wiki

Use this page as the object card for the model or method. The source page carries the evidence details, limitations, and agenda mapping.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in generative-recursive-reasoning-2026 rather than duplicating verdict rows here. At the entity level, GRAM is most relevant as an adjacent dynamic-compute and multi-hypothesis latent-state mechanism: it suggests a width-based alternative to only increasing recursive depth, but current evidence is puzzle-focused rather than time-series or action-conditioned world-model evidence.

Evidence

• Generative Recursive Reasoning

Official Artifacts

• Official project page: GRAM: Generative Recursive Reasoning
• Preprint: arXiv 2605.19376
• Workshop version: ICLR 2026 Workshop RSI poster
• Code status: official project page says code is coming soon; no official repository was verified at ingest time.

Related Pages

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