Generative Recursive Reasoning

Researchers introduced GRAM, a latent-variable framework for recursive reasoning that turns inference into stochastic multi-trajectory computation.

GRAM supports multiple solution paths, conditional and unconditional generation, and is trained with amortized variational inference.

It outperformed deterministic recurrent and recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks.

The approach offers a more flexible and scalable alternative for hard reasoning problems by enabling parallel multi-hypothesis reasoning.