Generative Recursive Reasoning

TL;DR

The paper introduces GRAM, a probabilistic recursive reasoning framework that models multiple hypotheses and solution strategies, enhancing reasoning and generation capabilities over deterministic models.

Contribution

GRAM extends recursive reasoning models into a probabilistic multi-trajectory framework, enabling multi-hypothesis reasoning and scalable inference.

Findings

GRAM outperforms deterministic baselines on structured reasoning tasks.

GRAM supports both conditional reasoning and unconditional generation.

It demonstrates improved inference scalability through recursive depth and parallel sampling.

Abstract

How should future neural reasoning systems implement extended computation? Recursive Reasoning Models (RRMs) offer a promising alternative to autoregressive sequence extension by performing iterative latent-state refinement with shared transition functions. Yet existing RRMs are largely deterministic, following a single latent trajectory and converging to a single prediction. We introduce Generative Recursive reAsoning Models (GRAM), a framework that turns recursive latent reasoning into probabilistic multi-trajectory computation. GRAM models reasoning as a stochastic latent trajectory, enabling multiple hypotheses, alternative solution strategies, and inference-time scaling through both recursive depth and parallel trajectory sampling. This yields a latent-variable generative model supporting conditional reasoning via $p_\theta(y \mid x)$ and, with fixed or absent inputs, unconditional generation via $p_\theta(x)$. Trained with amortized variational inference, GRAM improves over deterministic recurrent and recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks, while demonstrating an unconditional generation capability. https://ahn-ml.github.io/gram-website