Y Combinator · Video summary

Beyond Bigger Models: Recursion As The Next Scaling Law In AI

In this episode of Decoded, Y Combinator visiting partner Francois Shaard discusses the innovative use of recursion in AI models, specifically through Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM), which enhance reasoning capabilities without simply increasing model size. The conversation highlights how these models leverage recursive processes to solve complex problems more efficiently than traditional methods.

Y Combinator · May 1, 2026 · 38-min episode · 6-min read

Beyond Bigger Models: Recursion As The Next Scaling Law In AI

Key takeaways

The use of recursion in AI models like Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM) enhances reasoning capabilities without just increasing model size.
HRM achieved state-of-the-art performance on the ARC prize with only 27 million parameters, outperforming larger models like GPT-3 on specific tasks.
Traditional models face vanishing gradient problems as they scale, while recursive models maintain performance by reusing weights and employing fixed point iteration techniques.
Recursive models can handle incompressible problems more effectively, such as Sudoku and sorting algorithms, which traditional LLMs struggle with due to their architecture.
The backpropagation technique used in HRM allows for efficient training without the need to backpropagate through all recursion steps, minimizing computational overhead.
Explore the implementation of recursive models in your AI projects to enhance reasoning capabilities without increasing complexity.

Chapters

0:00🎯 Introduction to recursion in AI
0:11🎯 Historical context of RNNs and limitations
2:31🎯 Transformer models and their advantages
4:02🎯 Reasoning capabilities and limitations of LLMs
5:46🎯 Incompressible problems and external memory
7:40🎯 High-level overview of HRM (Hierarchal Reasoning Model)
9:55🎯 Key results and small model efficiency
12:06🎯 Optimization techniques in HRM
14:23🎯 Simplification and improvements over HRM
25:02🎯 Deep fixed point iteration and TRM recursion
34:13🎯 Potential of combining recursion with large models
35:36🎯 Summary and future outlook
36:22🎯 Different task-specific vs. general-purpose models

The digest

High-Level Overview

This discussion explores a promising direction in AI research: improving reasoning abilities by incorporating recursion during inference, rather than simply increasing model size. Two seminal 2025 papers introduce hierarchical and recursive architectures—Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM)—that outperform much larger models on complex tasks by leveraging recursion as a core computational principle. This shift challenges the dominant scaling paradigm focused solely on massive language models (LLMs), offering a more efficient path to advanced reasoning for researchers and practitioners interested in deep learning model design and AI capabilities.