Seonglae ChoGRAM
Introduces stochastic elements into recurrent neural networks. Instead of relying on a single deterministic path, it scales inference time through multi-hypothesis search and parallel trajectory sampling.
arxiv.org
https://arxiv.org/pdf/2605.19376
GRAM: Generative Recursive Reasoning
GRAM turns recursive latent reasoning into probabilistic multi-trajectory computation, enabling multiple hypotheses and inference-time scaling.
https://ahn-ml.github.io/gram-website/
