Generative Recursive Reasoning (GRAM)

I'm Matt Wood, and this is For Your Information. A live list of riffs and links for you and your agent, drawn from what I'm reading, noticing, questioning, concluding, and revising.

Core idea: Recursive Reasoning Models (RRMs) use repeated computation to refine a persistent latent state rather than append new elements to an output sequence. This decouples reasoning depth from both parameter scale and output length: a compact model can perform many steps of internal computation by repeatedly applying shared transition functions

arxiv.org →

Connections

Supports

MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling GRAM's recursive refinement of a persistent latent state offers a reasoning mechanism that could underlie the population-level test-time scaling MaxProof uses for generating and verifying proofs.

Neuronpedia Jacobian Lens: Interactive J-Space Explorer for Qwen3.6-27B Interpretability tooling like the Jacobian Lens could be applied to probe the persistent latent states that GRAM's recursive reasoning models rely on.

Type:	link
First published:	Thursday, 28 May 2026 at 12:00 UTC
Location:	https://arxiv.org/html/2605.19376v1
URL:	https://mattwood.fyi/i/cd960c1c