FROM THE FRONTIER - via SuperHuman
LLMs have a big problem. Unlike humans, who can reason internally without verbalizing every thought, LLMs are forced to express all reasoning as text (aka chain-of-thought prompting). This approach is inefficient — it requires copious amounts of training data and can fail if steps are slightly out of order. The result: AI that’s data-hungry, computationally expensive, and still falls short on complex reasoning.
A startup from Singapore set out to fix this bottleneck. Sapient Intelligence has unveiled the Hierarchical Reasoning Model (HRM), an AI architecture that beats traditional LLMs on complex reasoning tasks with a fraction of the data and compute.
Sapient’s HRM architecture outperforms traditional models on complex reasoning tasks. Source: arXiv
Here’s how it works: HRM mimics how human brains think. It uses 2 interconnected modules that work together like a supervisor and a worker. The worker module crunches the numbers on a sub-task until it reaches a solution. Then, the supervisor evaluates this progress, updates the overall strategy, and assigns the worker a new sub-task. This back-and-forth lets the AI think without verbalizing each step as text, requiring a fraction of the compute that goes into traditional models.
This could have significant implications:
-
David beats Goliath: HRM blew much larger AI models out of the water on complex reasoning tasks with 100x faster processing, achieving near-perfect scores on challenges where state-of-the-art models completely failed.
-
Real-world efficiency: For enterprises, this translates to major cost savings and new capabilities, with training times slashed to hours instead of months.
-
New frontiers: This approach shows promise in fields that require complex decision-making, like healthcare, climate forecasting, and robotics.