# how this tiny model beat ChatGPT on the “AGI” benchmark [HRM & TRM]

Source: https://www.youtube.com/watch?v=ZgwHaI2C-9s
Recap page: https://rapidrecap.app/video/ZgwHaI2C-9s
Generated: 2025-12-04T17:45:35.754+00:00

---
## Quick Overview

The Tiny Recursion Model (TRM) significantly outperforms the larger Hierarchical Reasoning Model (HRM) and surpasses current state-of-the-art Large Language Models (LLMs) on the ARC-AGI benchmark by employing a simpler, recursive, attention-free architecture with only 7 million parameters, achieving 40% accuracy on ARC-AGI-1 for only $1.76 per task.

**Key Points:**
- The Tiny Recursion Model (TRM), with only 7 million parameters, achieved 40% accuracy on ARC-AGI-1 and 6.2% on ARC-AGI-2, significantly beating larger LLMs like GPT-4 and Claude.
- TRM outperforms the Hierarchical Reasoning Model (HRM) on Sudoku-Extreme (87.4% vs 55.0%) and Maze-Hard (85.3% vs 74.5%), despite TRM being 4x smaller than HRM (7M vs 27M parameters).
- The HRM paper's initial success relied on heuristics like assuming the low-level network settled at equilibrium during training, which the TRM paper shows leads to inaccurate gradients.
- TRM achieves its performance without complex mathematical theorems, hierarchy, or biological arguments, relying on a single, tiny network structure that recursively updates a scratchpad.
- The authors of TRM justify its design by showing that increasing layers in HRM architecture only leads to overfitting on scarce training data, indicating that scaling up is not always the answer.
- TRM achieves its results using only MLPs in its blocks (instead of attention layers) and a single forward pass, which is simpler and more efficient than the two-module, multi-timescale approach of HRM.

![Screenshot at 0:04: The video contrasts the prevalent scaling paradigm, where AI companies relentlessly pursue larger LLMs \(Peter Griffin forcing more parameters onto LLMs\), against the emerging idea that progress is not solely dependent on parameter count, setting up the discussion about smaller, more efficient models.](https://ss.rapidrecap.app/screens/ZgwHaI2C-9s/00-00-04.png)

**Context:** This video analyzes the research behind the Tiny Recursion Model (TRM), a novel, small-scale AI architecture, contrasting it with the previous state-of-the-art, the Hierarchical Reasoning Model (HRM), particularly regarding performance on the Abstract and Reasoning Corpus (ARC) benchmark, often considered a test for Artificial General Intelligence (AGI). The video critiques the dominant scaling laws in AI, suggesting that for complex reasoning tasks with limited data, architectural simplicity and efficiency, as demonstrated by TRM, can surpass massive parameter counts.

## Detailed Analysis

The video argues that the trend of simply scaling up Large Language Models (LLMs) is flawed, especially for complex reasoning tasks like those in the ARC benchmark, as highlighted by the success of the Tiny Recursion Model (TRM). The TRM, developed by Alexia Jolicoeur-Martineau, uses only 7 million parameters and an MLP-based recursive architecture, achieving 40% on ARC-AGI-1 and 6.2% on ARC-AGI-2, outperforming much larger models and even beating the 27M parameter HRM on challenging tasks like Sudoku-Extreme (87.4% vs 55.0%). The HRM's approach, which mimics biological temporal separation with slow (high-level) and fast (low-level) recurrent networks, relies on the faulty heuristic that the fast network settles to equilibrium during training, leading to inaccurate gradient computations. TRM avoids this by using a single, tiny network that recursively updates a scratchpad (Z) and a canvas (Y) in a single forward pass, requiring no complex biological justifications or mathematical proofs like HRM. Furthermore, experiments show that increasing layers in the HRM architecture leads to overfitting when training data is scarce, reinforcing the 'Less is More' principle demonstrated by TRM's superior generalization with fewer parameters.

### The Scaling Debate

- More Parameters vs. Better Architecture: AI companies prioritize scale, burning up to $50 billion a year to build massive LLMs, but the video suggests scaling alone is insufficient for reasoning tasks.

### Hierarchical Reasoning Model (HRM)

- HRM uses two recurrent networks (high-level/slow, low-level/fast) inspired by brain temporal separation (theta and gamma waves) to solve sequential reasoning tasks.

### TRM

- The Tiny Recursion Model requires no complex theorems or biological arguments; it uses only a single tiny network (7M parameters) and MLPs (no attention) to recursively refine an answer via a scratchpad (Z) and canvas (Y).

### Empirical Results

- TRM achieves 40% on ARC-AGI-1 and 6.2% on ARC-AGI-2, beating HRM (27M parameters) on Sudoku-Extreme (87.4% vs 55.0%) and Maze-Hard (85.3% vs 74.5%).

### Critique of HRM Training

- HRM's training relies on assuming the low-level loop settles at equilibrium to compute gradients, but this assumption is often false, leading to inaccurate gradients and suboptimal learning.

### Lesson from Scaling Ablations

- Increasing layers in the HRM architecture led to decreased generalization due to overfitting when training data was scarce, proving that complexity does not always equate to better performance.

![Screenshot at 0:01: A Family Guy meme illustrates the industry consensus: AI companies believe bigger is better, forcing more parameters onto LLMs, contrasting with the video's theme.](https://ss.rapidrecap.app/screens/ZgwHaI2C-9s/00-00-01.png)
![Screenshot at 0:04: Sam Altman's quote about being willing to burn $50 billion a year to achieve AGI highlights the current industry focus on massive scale.](https://ss.rapidrecap.app/screens/ZgwHaI2C-9s/00-00-04.png)
![Screenshot at 0:19: The 'Change My Mind' meme asserts that 'Progress is not equal to Parameter Count,' challenging the industry's scaling-first mentality.](https://ss.rapidrecap.app/screens/ZgwHaI2C-9s/00-00-19.png)
![Screenshot at 0:36: A meme comparing the tiny 27M-parameter HRM against a trillion-parameter LLM emphasizes the surprising performance of smaller models.](https://ss.rapidrecap.app/screens/ZgwHaI2C-9s/00-00-36.png)
![Screenshot at 0:48: A bar chart comparing HRM \(blue\) against various LLMs \(grey/yellow\) on ARC-AGI and Sudoku/Maze benchmarks, showing HRM's superior performance for its size.](https://ss.rapidrecap.app/screens/ZgwHaI2C-9s/00-00-48.png)