A joint team from KAIST, Mila, and New York University—including Turing Award winners Yoshua Bengio and Sungjin Ahn—published a paper on May 19th on arXiv introducing GRAM (Generative Recursive reAsoning Models). Existing recursive reasoning models such as HRM, TRM, and Recurrent Transformers are deterministic: given the same input, they always generate identical reasoning paths, effectively limiting the set of valid solutions to just one attractor. GRAM transforms recursion itself into stochastic trajectories within a latent space, optimized via amortized variational inference. This enables the model to sample multiple hypothetical reasoning paths in parallel for a given input, thereby expanding computational scaling from mere ‘depth’ to also ‘breadth’. Importantly, this framework supports both conditional reasoning (p(y|x)) and serves as an unconditional generative model (p(x)), allowing independent sampling of the distribution of reasoning problems themselves.
Despite having only 10 million parameters, GRAM outperformed similarly sized models across several challenging benchmarks: achieving 97.0% accuracy on Sudoku-Extreme (compared to 87.4% for TRM under the same conditions), over 90% coverage on N-Queens, 52.0% on ARC-AGI-1, and 11.1% on ARC-AGI-2—results competitive with much larger language models. According to the paper, this work was initially presented as a poster at the ICLR 2026 Workshop on Recursive and Structured Reasoning (RSI) in March 2026; the current arXiv version represents the full manuscript. Concurrently, HRM-Text—a 1-billion-parameter pre-trained model built on the same recursive architecture—was released this week. Together, these two studies provide empirical evidence supporting dual progress in recursive reasoning research: one line targeting discrete reasoning tasks with small parameter counts, the other extending such methods to large-scale language modeling contexts.