A model architecture that solves complex problems through structured, multi-level reasoning steps.
A Hierarchical Reasoning Model (HRM) is a neural network architecture designed to decompose complex reasoning tasks into structured, multi-level processes. Rather than attempting to solve a problem in a single forward pass or a flat chain of steps, an HRM organizes computation into distinct hierarchical layers or stages — typically a higher-level "thinking" layer that plans and coordinates, and one or more lower-level layers that execute specific sub-tasks. This separation of abstract planning from concrete execution mirrors how humans approach difficult problems: first forming a high-level strategy, then working through the details.
The core mechanism of an HRM involves recurrent or iterative processing across two or more levels of abstraction. A high-level module operates at a coarser temporal or conceptual granularity, generating latent representations that guide a lower-level module operating at finer resolution. These latent states are not necessarily human-interpretable tokens — they function as internal scratchpad representations that allow the model to "think" before producing output. This design allows the model to allocate more compute to harder problems dynamically, a property sometimes called adaptive computation, and enables longer chains of reasoning without requiring proportionally longer output sequences.
HRMs are particularly relevant in the context of scaling test-time compute, where the goal is to improve model performance not by increasing parameter count but by allowing more sophisticated reasoning at inference time. By structuring reasoning hierarchically, HRMs can avoid some of the pitfalls of flat chain-of-thought approaches — such as error propagation through long token sequences — while maintaining interpretability advantages over purely implicit reasoning. They have shown promise on tasks requiring multi-step mathematical reasoning, algorithmic problem solving, and planning under constraints.
The significance of HRMs lies in their potential to bridge the gap between the raw pattern-matching capabilities of large language models and the structured, deliberate reasoning associated with human cognition. As AI systems are increasingly deployed in high-stakes domains requiring reliable multi-step inference, architectures that explicitly encode hierarchical structure offer a principled path toward more robust and controllable reasoning. HRMs represent an active area of research at the intersection of cognitive science-inspired design and modern deep learning.