Gram: Recursive reasoning models with stochastic latent trajectories (10M param) (arxiv.org)

Researchers have introduced Generative Recursive Reasoning Models (GRAM), a novel framework that enhances the capabilities of neural reasoning systems by transforming deterministic recursive reasoning into a probabilistic, multi-trajectory approach. Unlike traditional models that follow a single latent trajectory leading to one definitive prediction, GRAM leverages stochastic latent trajectories to explore multiple hypotheses and potential solutions simultaneously. This innovation not only allows for greater flexibility and scalability during inference but also enables conditional reasoning based on provided inputs and unconditional generation when inputs are fixed or absent. The significance of GRAM lies in its potential to advance structured reasoning tasks and multi-solution constraint satisfaction challenges, outperforming existing deterministic recursive models. By utilizing amortized variational inference, GRAM provides improved results in deploying generative latent-variable models, positioning itself as a critical development in the AI/ML landscape. Its ability to support both recursive depth and parallel sampling enhances its application scope, making it a promising tool for future AI systems focused on complex reasoning and generative tasks.