Hierarchical Reasoning Model

The Problem: Reasoning Is Expensive

Modern large language models (LLMs) typically solve complex reasoning tasks through Chain-of-Thought (CoT): the model writes out its thinking process explicitly as text, step by step. This works — but it is expensive. More tokens, more compute, more cost.

On top of that, more complex tasks typically require more parameters. GPT-4 is estimated at roughly 1.8 trillion parameters. For many businesses, that is not a realistic option — neither in terms of cost nor data privacy, when data must not leave the premises.

The HRM paper asks a fundamentally different question: does reasoning really need to be that large and that explicit?

The Approach: Hierarchical Recurrent Processing

The core idea of HRM is simple but powerful: reasoning happens on two levels simultaneously — similar to how humans plan strategically and execute concrete actions without narrating every thought out loud.

The model is recurrent: it processes the same input multiple times across several "thinking rounds" (recurrent steps) before producing an answer. This internal processing space replaces the explicit reasoning text used in CoT.

• No Chain-of-Thought required: Thinking happens internally in the network's activations, not as visible text.
• Only 1000 training examples: Instead of millions of samples, a tiny dataset suffices — a remarkable result.
• 27 million parameters: Small enough to run locally on consumer hardware.

The Architecture: Two Modules, Two Levels

HRM consists of two recurrent modules that work together hierarchically:

• High-Level Module: Responsible for abstract, strategic planning. It processes the task at a high level and provides direction and strategy. This module runs slowly — few iterations per task.
• Low-Level Module: Handles detailed, concrete computation. It carries out the strategy set by the high-level module in small steps. This module runs faster and iterates more frequently.

Both modules exchange state information — the high-level module can adjust its state when the low-level module encounters obstacles. This creates a dynamic loop between planning and execution.

Results: Small Beats Large

The authors test HRM on three different task types that require different kinds of reasoning:

• Sudoku: The model reliably solves Sudoku puzzles. Sudoku requires systematic elimination and backtracking — classic strengths of hierarchical planning.
• Maze Navigation: HRM finds paths through complex mazes. The high-level module plans the overall route, the low-level module navigates the individual steps.
• ARC (Abstraction and Reasoning Corpus): ARC is considered a particularly hard benchmark for abstract pattern recognition. HRM achieves notable results — without any Chain-of-Thought.

The remarkable part: these results are achieved with only 27 million parameters and 1000 training examples — a fraction of what comparable approaches require.

What Does This Mean for My Business?

The HRM paper is especially relevant for SMEs that want or need to run AI locally — for example due to data privacy requirements (GDPR) or limited budgets. Three concrete implications:

• Local deployment becomes more realistic: A 27M parameter model runs on any modern laptop or small server. No cloud dependency, no data sharing with external providers.
• Specialized small models: Research shows that for clearly defined tasks (planning, routing, puzzles, structured decisions), small specialized models can perform surprisingly well — at a fraction of the cost.
• Fine-tuning with few data points: 1000 training examples is realistic for many businesses. This means your own data could be enough to train a specialized model without needing to collect massive datasets.

Important caveat: HRM is not a universal model. It is optimized for structured reasoning tasks, not free text generation or conversation. The use case makes sense where clear rules and structured problems dominate — for example in planning systems, process optimization, or decision support.

Context: Why Does This Paper Matter?

The HRM paper appeared in June 2025 and sits within a growing research direction questioning whether large models are really necessary for everything. While industry pushes toward ever-larger models, this research shows that architecture can replace size.

The idea of splitting reasoning into two levels — strategic and tactical — is not new. It appears in classical AI planning (STRIPS, HTN), cognitive science (System 1 and System 2 per Kahneman), and robotics. HRM successfully transfers this principle into a neural network.

For practitioners, this means: the search for the right model should not start only with size and benchmark scores, but with architecture. A well-designed small model can outperform a generalist large model on specific tasks — while being faster and cheaper.