# Hierarchical Reasoning Models (HRMs), Simply Explained #65 Aug 25, 2025 LLMs https://www.aibutsimple.com/p/hierarchical-reasoning-models-hrms-simply-explained In June 2025, a [research paper](https://arxiv.org/pdf/2506.21734?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained) published by Guan Wang and [Sapient Intelligence’s](https://www.sapient.inc/?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained) team rocked the AI world, showcasing a highly powerful LLM model that blew concurrent reasoning models out of the water. This new LLM architecture, called the Hierarchical Reasoning Model (HRM), promises extraordinary accuracy along with efficiency and stability, challenging our existing ideas of training models and receiving outputs. ![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,width=1200,format=auto,onerror=redirect,quality=80/uploads/asset/file/b1c3f52d-a400-4917-a1fb-7d0a635d483d/Copy_of_AI__but_Simple__1024_x_400_px_.png) The release went viral within the AI community, with X/Twitter discussions hitting over 4 million views and tens of thousands of likes, while YouTube videos dissecting the work surpassed 475K views. But unlike typical AI breakthroughs that require massive computational resources and internet-scale datasets, HRM achieves its remarkable results with just 27 million parameters and 1,000 training examples. - This sounds like a lot to the untrained eye, but consider that models like ChatGPT and Llama rely on billions of parameters and huge training data sets. With the paper only coming out two months ago, there isn’t much supporting evidence, apart from what’s included in the paper, to explore. For that reason, let’s take a broader view of how HRM works to gain a broader understanding of why the community is viewing it as a potential paradigm shift. --- ### Start learning AI in 2025 Keeping up with AI is hard – we get it! That’s why over 1M professionals read [Superhuman AI](https://www.superhuman.ai/c/confirmation?magiclink_subscription) to stay ahead. - Get daily AI news, tools, and tutorials - Learn new AI skills you can use at work in 3 mins a day - Become 10X more productive [Sign up and start mastering AI](https://www.superhuman.ai/c/confirmation?magiclink_subscription) --- ### What is a Hierarchical Reasoning Model? Today’s AI landscape is focused on building models that can solve complex challenges, which usually require multiple modes of thought and levels of problem-solving. Current large language models primarily rely on Chain-of-Thought (CoT) prompting, breaking down problems into intermediate text-based steps, essentially forcing the model to "think out loud" as it works toward a solution. The problem with CoT is that while it lends LLMs the ability to work “piece by piece,” it demands huge amounts of training data and results in slow outputs. Additionally, CoT-aided LLMs still struggle to work through puzzles efficiently, as they can only operate within their available context window, which is a pre-allocated working memory. CoT can also be problematic if the model makes a mistake at any point. Just like a student attempting a math problem and messing up at some point, the model will likely not catch its mistake if it isn’t immediately apparent. According to researchers at Sapient Intelligence, "CoT for reasoning is a crutch, not a satisfactory solution." ![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,width=1200,format=auto,onerror=redirect,quality=80/uploads/asset/file/71d26ba6-9cd3-4f15-a8e8-0c9b3c2cd338/unnamed.png) More fundamentally, current AI models are architecturally "shallow." LLMs always perform a predefined quantity of computation to produce each output token. Most models won’t spend additional resources exploring alternative avenues, locking them into one line of thinking, which has a possibility to be incorrect. They're locked into their reasoning depth from the moment they've been prompted, unable to dynamically adjust based on problem complexity. ![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,width=1200,format=auto,onerror=redirect,quality=80/uploads/asset/file/bce06308-7b94-46a0-b076-507e26abf621/Copy_of_AI__but_Simple__1024_x_500_px___1_.png) ### How do Hierarchical Reasoning Models work? Seeking a better approach, the Sapient team turned to neuroscience for a solution, noting that "the human brain provides a compelling blueprint for achieving the effective computational depth that contemporary artificial models lack." Now this sounds like a direct pull from science fiction, but it is important to remember: the brain is still the most powerful processing engine we currently know of. The brain naturally organizes any challenge it faces and breaks it down while simultaneously delegating different cortical regions to solve it. It’s this superpower that enables deep, multi-stage reasoning that current AI architectures simply cannot replicate. HRM mirrors this biological design through two collaborative recurrent modules that work in harmony: ![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,width=1200,format=auto,onerror=redirect,quality=80/uploads/asset/file/83931b04-4ebc-4fe2-a237-eb60cebd35a4/Copy_of_AI__but_Simple__1024_x_550_px___1_.png) The High-Level (H) Module: The H Module approaches the problem from the macroscopic “bird’s eye view,” working on the overall problem-solving strategy and long-term planning. Think of it as the executive making high-level decisions about how to approach their next move in the market. The Low-Level (L) Module: Conversely, the L Module works within the microscopic level, handling detailed computations and intensive resource expenditure; it’s the analyst crunching the numbers in the cubicles outside the executive’s office. Just as the left and right sides of your brain are responsible for logical decisions and creative thinking, respectively, these modules work together through what researchers call "hierarchical convergence." The fast L-module addresses a portion of the problem, executing multiple steps until it reaches a stable solution. At that point, the slow H-module takes this result, updates its overall strategy, and gives the L-module a new, refined sub-problem to work on. We discussed a similar looping concept in [Issue #59’s discussion of Agentic AI](https://www.aibutsimple.com/p/agentic-ai-simply-explained?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained). This process is different from CoT because the model now takes care of these steps in the background, in a latent space. All this means is that the model doesn’t express its steps in text form, but rather keeps them internally as numerical data. - This aligns with scientific understanding that "the brain sustains lengthy, coherent chains of reasoning with remarkable efficiency in a latent space, without constant translation back to natural language." When you solve a complex problem, you probably don’t take time to write out each micro-decision, which is mirrored in the architecture of these HRMs. This internal reasoning happens through collaboration between the H and L modules, where the slow-updating H module advances only after the fast-updating L module has completed multiple computational steps and reached a local equilibrium, or in other words, a place to halt before proceeding. ![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,width=1200,format=auto,onerror=redirect,quality=80/uploads/asset/file/46ebcfd0-ca9e-43eb-beb3-34296080f664/Copy_of_AI__but_Simple.png) When analyzing HRMs internal processes, the researchers were astounded: the model appears to adapt different reasoning strategies for different types of problems. In maze navigation, it initially explores multiple potential paths simultaneously, then eliminates blocked routes and constructs the final solution. For Sudoku puzzles, it resembles humans, exploring potential solutions and backtracking when hitting dead ends, similar to how we solve them in the newspaper. On [ARC](https://arcprize.org/?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained) (an AI benchmark for intelligence) tasks, it makes incremental adjustments, iteratively improving solutions until it reaches the optimal one. ### The Results The numbers speak for themselves. On Sudoku-Extreme puzzles, HRM solved 55% of them, while Claude 3.7 and OpenAI's o3-mini-high failed completely. On 30×30 mazes, HRM found the optimal path 74.5% of the time, while other models again scored 0%. On the ARC-AGI benchmark (developed to be likened to an IQ test for AI), HRM scored 40.3% using only 1,000 training examples. - Again, it may look unremarkable, but context matters: other leading CoT-based models like o3-mini-high scored a 34.5%, and Claude 3.7 8K context came in at 21.2%, despite their considerably larger parameter numbers and context lengths. HRM's design is so lean and flexible that Guan Wang said, perhaps slightly exaggerated, that it can be trained to solve pro-level Sudoku in just two GPU hours. Compare this to the months of training and massive computational resources required for today's largest language models. --- ### [Level Up Your Career With AI, But Simple PRO](https://www.aibutsimple.com/upgrade?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained) ![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,width=1200,format=auto,onerror=redirect,quality=80/uploads/asset/file/281697a8-c915-435c-a68b-43374b671698/aibutsimple_academy.png) Supporting us by purchasing a PRO membership will get you: - Exclusive Hands-on Tutorials With Code - Full Math Explanations and Beautiful Visualizations - Ad-Free Reading for an Uninterrupted Learning Experience [Break into AI or sharpen your edge by joining today!](https://www.aibutsimple.com/upgrade?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained) --- ### Limitations and Challenges Despite its impressive performance, HRM faces several limitations that users and organizations must consider. The architecture's reasoning process occurs in latent space, creating interpretability challenges. Unlike CoT models, where you can read the "thinking" tokens, HRM's internal reasoning is opaque, similar to “black box” neural networks. When independently verified by the ARC Prize organization, HRM scored 2% on the ARC-AGI-2 benchmark, the more difficult, older brother of the first one, which isn't considered meaningful progress from 0%. This suggests that while HRM excels at certain types of reasoning tasks, it may not generalize across all reasoning domains. The model also requires a precise tuning of its hierarchical convergence parameters, which dictate the H and L modules. The training process, while efficient, requires expertise in managing the interaction between the H and L modules. Most importantly, general skepticism still remains. While the researchers did publish their code base, many still are concerned with the fact that these inflated scores could be a result of overfitting, where a model can score extremely highly under specific conditions but can crash anywhere else. ![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,width=1200,format=auto,onerror=redirect,quality=80/uploads/asset/file/d3054899-2e00-4f8b-bdc8-c96c52558fc1/Copy_of_AI__but_Simple__1024_x_400_px___1_.png) Other critics remain focused on the fact that these models have only been tested on game mechanics and tests, whereas the majority of everyday users would be interested in more general applications. Lastly, the problem that plagues these findings is replicability. With its groundbreaking status, unless external labs and institutes are able to validate Sapient Intelligence’s findings, HRMs seem to still be up in the air. ### The Bottom Line HRM’s early-stage performance suggests that brain-inspired designs could unlock powerful reasoning at a fraction of the cost, an important consideration as current upgrades in models also mean sky-high resource spending. This is setting the stage for alternative models, including [Sakana's continuous thought machines](https://sakana.ai/ctm/?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained), 1-bit LLM "bitnets," and even diffusion-based models that Google is experimenting with. The success of HRM raises key questions about how to improve models. Instead of simply making models bigger and training them on more data, perhaps the future lies in smarter architectural designs that mimic the computational principles that natural evolution has refined over billions of years. After all, “Brains over brawn.” --- Here’s a special thanks to our [biggest supporters](https://www.aibutsimple.com/upgrade?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained): - Sushant Waidande - Sai Krishna Pashapu --- If you enjoy our content, consider supporting us so we can keep doing what we do. Please share this with a friend! Want to reach 50000+ ML engineers? Let’s work together: [WORK WITH US 📈](https://www.passionfroot.me/aibutsimple?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained) Feedback, inquiries? Send us an email at [edwin@aibutsimple.com](mailto:edwin@aibutsimple.com). If you like, you can also donate to our team to push out better newsletters every week! [Support us ☕](https://ko-fi.com/aibutsimple?utm_source=newsletter.aibutsimple.com&utm_medium=referral&utm_campaign=hierarchical-reasoning-models-hrms-simply-explained) --- That’s it for this week’s issue of AI, but simple. See you next week! —AI, but simple team