Less is More: Recursive Reasoning with Tiny Networks (Paper Review)

Non-members can read for free
Modern AI often chases scale: deeper layers, more attention heads, and billions of parameters. But hidden beneath this race lies a quieter revolution: recursive reasoning, the idea that a model can improve its own thoughts, not by growing larger, but by thinking again.

This is where the story of Hierarchical Reasoning Models (HRM) and Tiny Recursion Models (TRM) unfolds: two architectures that dared to ask whether depth itself could be simulated, rather than built.

The Roots of Recursive Reasoning

Recursion is not just a mathematical trick, it’s how understanding evolves.
A model that reasons recursively doesn’t generate an answer once; it revisits its reasoning steps, refining what it already knows. Each iteration becomes a reflection, or a quiet dialogue between past and present states.

But recursion comes with a paradox:
without control, it becomes endless.
With too much control, it becomes shallow.

The challenge, then, is not to make recursion deeper but to make it aware.

Where Hierarchical Reasoning Models (HRM) started