Member-only story
Tiny Recursive Models: Achieving Better Reasoning with Radical Simplicity
A 27-million parameter model beat the hardest AI benchmarks using brain-inspired hierarchical reasoning. Researchers celebrated the breakthrough. Then someone asked: Where are the ablations?
Hierarchical Reasoning Models (HRM) seemed revolutionary — two coordinated models that mimic human dual-system thinking, adaptive computation budgets, and reinforcement learning for stopping criteria. Complex, sophisticated, biological.
Then Tiny Recursive Models (TRM) systematically removed each “essential” component and discovered something embarrassing: nearly all of it was unnecessary.
This comparison reveals which architectural choices matter for reasoning so far and which add unnecessary complexity. More importantly, it highlights Deep Supervision as the critical component driving performance, not the hierarchical structure.
Understanding the HRM Foundation
HRM draws inspiration from dual-process theory in cognitive psychology.
Humans employ two decision-making systems: a fast, intuitive “System 1” for routine tasks, and a slower, deliberative “System 2” for complex reasoning.
HRM translates this into a machine learning architecture with two distinct…
