Guan Wang

The HRM-Text paper is now available 🎉 HRM-Text explores a different approach to language model pretraining: hierarchical recurrent computation, task-completion training, and latent-space reasoning. At just 1B parameters, HRM-Text achieves competitive performance with dramatically lower training cost and data requirements. 1B parameters 40B unique tokens ~1 day of pretraining ~$1000 training cost

Huge congrats on reproducing the HRM-Text results! 🚀Pretraining an LLM on just 16 GPUs is a massive win. 👏#AI #NeuroAI #LLM #HRMText

Congratulations to the team! 👏Excited to see new explorations building on the HRM architecture and pushing recursive reasoning further.

Paper: arxiv.org/abs/2605.20613 GitHub: github.com/sapientinc/HRM… Hugging Face: huggingface.co/sapientinc/HRM…

The HRM-Text paper is now available 🎉 HRM-Text explores a different approach to language model pretraining: hierarchical recurrent computation, task-completion training, and latent-space reasoning. At just 1B parameters, HRM-Text achieves competitive performance with dramatically lower training cost and data requirements. 1B parameters 40B unique tokens ~1 day of pretraining ~$1000 training cost

In this benchmark deep-dive, Sapient’s founders William and Guan are joined by research team members Changling and Yasin to unpack HRM-Text’s performance across MATH, DROP, ARC-Challenge, and MMLU. 📊 Beyond the scores, they discuss what each benchmark measures, how HRM-Text compares with larger models, and why efficiency matters. Watch the full discussion to learn more about HRM-Text and Sapient’s leaner path toward general intelligence.

HRM-Text 101 is here. This tutorial takes you from zero to one: from setup to fine-tuning to evaluation. Download the base checkpoint. Fine-tune it on a real task. Evaluate the results. End to end, on a single GPU. Watch the tutorial and start building with HRM-Text.

Introducing HRM-Text. An ultra-lean 1B-parameter reasoning language model designed to deliver strong general performance with a fraction of the data, compute, and infrastructure. Trained on just 40B structured tokens, HRM-Text achieves competitive performance while using ~1/1000 of the training data of comparable models. The kicker? The full model trains in roughly one day on a $1,000 budget. This opens the door to a new generation of AI that is powerful, accessible, and radically easier to adapt. Theories and research concepts once deemed too expensive to test are officially back in the game. Sapient Intelligence invites you to help us shape a new paradigm for general intelligence.

Tomorrow, we will unveil a new path to general intelligence. Lean. Powerful. Efficient. The countdown is on⏳

Behind the code, there is a specific kind of expertise. We are a team of researchers and engineers rooted in the labs of Tsinghua University, University of Cambridge, University of Alberta, Carnegie Mellon University, and Peking University—with experience at DeepMind, DeepSeek, xAI, and more. We've seen the limits of the current AI architectures firsthand from within the organizations that scaled them. Now, across three countries, we are building an alternative. We aren't just shipping another wrapper; we are shipping a new fundamental architecture.

Hierarchical reasoning works well on large language models!🎉

🔥It’s official-Sapient HRM Discord Community is now live！ This is a place to discuss, connect, and collaborate as we shape HRM’s future together. We will be sharing our latest work, releases, and tips, as well as hosting Q&A sessions💬💬 Hop on this journey with us as we push the boundaries of what HRM and AGI at large can achieve！🙌 ➡️Join us on Discord here discord.gg/sapient

Thanks to @arcprize for reproducing and verifying the results! ARC-AGI-1: public 41% pass@2 - semi private 32% pass@2 ARC-AGI-2: public 4% pass@2 - semi private 2% pass@2 Due to differences in testing environments, a certain amount of variance in results is acceptable. According to tests run on our infrastructure, the open-source version of HRM on our GitHub can achieve a score of 5.4% pass@2 on the ARC-AGI-2. We welcome everyone to run it on your own infra and share your scores~ This is our first submission to the leaderboard, and it's a good starting point. We appreciate everyone for your support and feedback on HRM, both before and after our appearance on the ARC leaderboard. All of this encourages and motivates us to improve. The hierarchical architecture is designed to resolve premature convergence in long-horizon tasks, like master-level Sudoku that takes hours for humans to solve. See the comparison with a simple recurrent Transformer. Such a long chain might not be essential for ARC problems, and we only used a high-low ratio of 1/2. Larger ratios are often needed for optimal performance for Sudoku problems. In the case of ARC-AGI, the success of HRM is a testament to the model's ability to exhibit fluid intelligence - that is, its capability to infer and apply abstract rules from independent and flat examples. We are glad it was discovered in a recent blog post that the outer loop and data augmentation are essential for this ability, and we especially thank @fchollet @GregKamradt @k_schuerholt for pointing this out. Finally, we are accelerating the iteration of the HRM model and continuously pushing its limits, with good progress so far. At the same time, we believe the hierarchical architecture is highly effective in many scenarios. Moving forward, we will make further targeted updates to the architecture and validate it on more applications. We will also release an FAQ to address the key questions raised by the community. 🧠 Stay tuned!

❤️ Thanks for highlighting our HRM paper! Apologies for any confusion—we're working on clarifying this thoroughly. Stay tuned for updates!

@taihongtran Search just finds stuff. HRM learns patterns and reasons. We’re testing if it can work as a search after training - stay tuned.

@makingAGI What is the differences compare to neural search algorithms?

🚀Introducing Hierarchical Reasoning Model🧠🤖 Inspired by brain's hierarchical processing, HRM delivers unprecedented reasoning power on complex tasks like ARC-AGI and expert-level Sudoku using just 1k examples, no pretraining or CoT! Unlock next AI breakthrough with neuroscience. 🌟 📄Paper: arxiv.org/abs/2506.21734 💻Code: github.com/sapientinc/HRM

@evilmathkid We have clarified it here: github.com/sapientinc/HRM…

@makingAGI @makingAGI The writing is a little unclear, could you please clarify which dataset are the ARC results on? Public train, public eval, private eval? Can't find you on the leaderboard on kaggle

@ADarmouni No. Eval data stayed separate. Details here: github.com/sapientinc/HRM…

@makingAGI Something I need clarification: are the Arc-AGI results on data in the training set? You mention public eval in train data, so I suppose you trained on it But is it also the data you evaluate it on? Even if it overfits the training data, results are super cool!

@Homo_Immortalis Answers will not be coming soon, but we’re working on it!

@makingAGI This is exciting Guan, when do you think AI will be able to answer humanity’s unsolvable questions, like reversing aging, or understanding consciousness?

@NewAgeNihilism Thanks! Any questions, email us anytime: research@sapient.inc.

@makingAGI Great job, really good paper. I'm still wrapping my head around it

@codeslubber Fair point! We’re inspired by brains, not handmade expert-system hierarchies. We let the hierarchy learn and grow from data, not hand-made rules.

Will read the paper, but saying that by making the code hierarchical we are using neuroscience is kind of silly no? I am super interested in how these hierarchies manifest. One of the reasons Expert Systems failed was there never was any consensus on how to organize knowledge representation, and encoding everything in a bespoke fashion was not scalable.