New AI model explores multiple reasoning paths in parallel

Who: Posted by Sungjin Ahn (@SungjinAhn_), a machine learning researcher at KAIST (Korea Advanced Institute of Science and Technology), sharing his own team's paper. Co-authors include Junyeob Baek, Mingyu Jo, and Minsu Kim (also at KAIST and Mila, the Montreal AI institute), along with Mengye Ren at NYU and Yoshua Bengio at Mila — Bengio being one of the most cited researchers in the history of deep learning.

What's new: The team introduces (Generative Recursive reAsoning Model), a new reasoning architecture that injects randomness into a class of models previously known for being purely deterministic. Where older like , , and always produce the same answer given the same input, GRAM can explore many different reasoning paths simultaneously.

How it works: Instead of following one fixed chain of thought, GRAM treats its internal reasoning steps as a . At each loop iteration, the model samples from a range of possible next-step hypotheses rather than committing to one. This gives GRAM two levers for scaling up its thinking: going deeper (more loops) or going wider (running many parallel sampled trajectories and comparing results). A notable bonus is that the same mathematical setup that lets GRAM answer questions also lets it function as a , capable of producing new data on its own.

The numbers: Despite having only 10 million , GRAM scores 97.0% on Sudoku-Extreme, compared to 87.4% for TRM. On , it reaches 52.0%, and 11.1% on the harder . It also solves over 90% of cases in the coverage benchmark.

Why it matters: The core insight — that adding controlled randomness to a looping reasoner lets it explore multiple hypotheses in parallel — is a meaningful structural departure from how most current reasoning models work. The fact that this is achieved with a very small model suggests the gains come from the architecture itself, not from simply throwing more computing power at the problem. The paper and project page are both public.