Recursion Is The Next Scaling Law In AI

Why It Matters

Recursive inference offers a cost‑effective route to higher reasoning ability, potentially redefining AI scaling strategies and accelerating progress toward more capable, energy‑efficient systems.

Key Takeaways

• Recursion at inference boosts reasoning without larger models.
• HRM uses hierarchical loops with shared weights for puzzle solving.
• TRM extends deep equilibrium methods to improve backpropagation.
• Small 27M‑parameter HRM outperformed larger models on ARC tasks.
• Recursive inference may become next scaling law for AI efficiency.

Summary

The Decoded episode spotlights recursion as the emerging scaling law in AI, focusing on two 2025 papers—Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM)—that demonstrate how repeated inference steps can boost reasoning performance without simply enlarging model size.

The hosts contrast classic RNNs, which suffer from back‑propagation‑through‑time and vanishing gradients, with modern transformers that avoid those issues but lack intrinsic temporal compression. They argue that LLMs hit a theoretical limit when solving incompressible tasks like sorting or Sudoku, because each token must be processed in a single forward pass limited by the number of layers.

HRM introduces three nested recursion levels—low‑frequency, high‑frequency, and outer refinement—using the same weights repeatedly. By applying a deep‑equilibrium (DEQ) style fixed‑point iteration, the model treats successive hidden‑state updates as a mini‑batch, sidestepping full BPTT. The 27‑million‑parameter HRM achieved roughly 70 % accuracy on ARC‑Challenge, surpassing larger pretrained baselines trained on only a thousand puzzle examples.

If recursive inference can consistently replace brute‑force scaling, AI developers could train far smaller models that still solve complex reasoning tasks, cutting compute costs and energy consumption. The approach also highlights the need for external memory mechanisms to break current algorithmic lower bounds, pointing to a new research frontier for efficient, brain‑inspired AI.

Original Description

A 7-million parameter model outperforming models a thousand times its size on tasks like ARC Prize. That's what recursive reasoning unlocks.

In this episode of Decoded, YC's Ankit Gupta and Francois Chaubard break down two recent papers on recursive AI models, HRMs and TRMs, that are achieving state-of-the-art results with a fraction of the parameters of today's largest models.

They explain why standard LLMs hit a fundamental ceiling on certain reasoning tasks, how recursion at inference time gives small models the compute depth to break through it, and what happens when you combine these ideas with the power of large-scale foundation models.

Apply to Y Combinator: https://www.ycombinator.com/apply

Work at a startup: https://www.ycombinator.com/jobs

00:00 - Intro

00:35 - Model Foundations

01:15 - RNN Limits and LLM Contrast

02:36 - Reasoning Limits and Sorting Analogy

04:22 - HRM Paper Introduction

05:25 - HRM Architecture and Intuition

07:36 - HRM Results and Outer Loop

09:46 - TRM Paper Overview

11:20 - TRM Training and Fixed Point

13:30 - Detailed HRM Summary

20:46 - Comparing HRM and TRM

34:45 - Future Outlook and Outro