GRAM accepted to NeurIPS 2026: recursive reasoning becomes stochastic latent trajectories

SungjinAhn_ · x · 2026-09-25

GRAM (Generative Recursive reAsoning Models) from KAIST, Mila, NYU and Yoshua Bengio is accepted to NeurIPS 2026.

Existing recursive reasoners (HRM, TRM, Looped Transformers) are deterministic: the same input always converges to one answer, collapsing all plausible reasoning paths into a single attractor.
GRAM models recursion as a stochastic latent trajectory, keeping multiple hypotheses and alternative strategies, with inference-time scaling in both depth and width (parallel trajectory sampling).
The same formulation supports conditional reasoning p(y|x) and unconditional generation p(x). Trained with amortized variational inference, it beats deterministic baselines on structured reasoning and multi-solution constraint satisfaction; revised paper and code coming soon.

More from Research