Generative Recursive Reasoning Model Introduced for Multi-Trajectory Computation
A new framework called Generative Recursive Reasoning Models (GRAM) is proposed in a preprint on arXiv (2605.19376v1). GRAM extends Recursive Reasoning Models (RRMs) by introducing probabilistic multi-trajectory computation, allowing for multiple hypotheses and alternative solution strategies. It models reasoning as a stochastic latent trajectory, enabling inference-time scaling through recursive depth and parallel trajectory sampling. GRAM functions as a latent-variable generative model supporting conditional reasoning via pθ(y|x) and unconditional generation via pθ(x). The work addresses limitations of existing deterministic RRMs that follow a single latent trajectory and converge to a single prediction. The paper is authored by researchers and published on arXiv under the title "Generative Recursive Reasoning".
Key facts
- New framework GRAM turns recursive latent reasoning into probabilistic multi-trajectory computation.
- GRAM extends Recursive Reasoning Models (RRMs) by introducing stochastic latent trajectories.
- It enables multiple hypotheses and alternative solution strategies.
- Inference-time scaling is achieved through recursive depth and parallel trajectory sampling.
- GRAM supports conditional reasoning via pθ(y|x) and unconditional generation via pθ(x).
- Existing RRMs are largely deterministic and converge to a single prediction.
- The paper is available on arXiv with ID 2605.19376v1.
- The title of the paper is 'Generative Recursive Reasoning'.
Entities
Institutions
- arXiv