Beyond LLM Reasoning Limits

Current LLMs depend heavily on Chain-of-Thought prompting, an approach that often suffers from brittle task decomposition, immense training data demands and high latency. Inspired by the hierarchical and multi-timescale processing in the human brain, HRM overcomes these constraints by embracing three fundamental principles observed in cortical computation: hierarchical processing, temporal separation, and recurrent connectivity. Composed of a high-level module performing slow, abstract planning and a low-level module executing rapid, detailed computations, HRM is capable of alternating dynamically between automatic thinking ("System 1") and deliberate reasoning ("System 2") in a single forward pass.

"AGI is really about giving machines human-level, and eventually beyond-human, intelligence. CoT lets the models imitate human reasoning by playing the odds, and it's only a workaround. At Sapient, we're starting from scratch with a brain-inspired architecture, because nature has already spent billions of years perfecting it. Our model actually thinks and reasons like a person, not just crunches probabilities to ace benchmarks. We believe it will reach, then surpass, human intelligence, and that's when the AGI conversation gets real," said Guan Wang, founder and CEO of Sapient Intelligence.

Benchmark Breakthroughs

Despite its compact scale of 27 million parameters and using only 1000 input-output examples,all without any pre-training or Chain-of-Thought supervision, HRM learns to solve problems thateven the most advanced LLMs struggle with. In the Abstraction and Reasoning Corpus (ARC) AGI Challenge, a widely accepted benchmark of inductive reasoning, HRM archives aperformance of 5% on ARC-AGI-2, significantly outperforming OpenAI o3-mini-high, DeepSeekR1, and Claude 3.7 8K, all of which rely on far larger sizes and context lengths. In complex Sudoku puzzles and optimal pathfinding in 30x30 mazes, where state-of-the-art CoT methods completely fail, HRM delivers near-perfect accuracy.

The Sapient Intelligence team is already running new experiments and expect to publish even stronger ARC-AGI scores soon.

Real-World Impact

HRM data efficiency and reasoning accuracy open new opportunities in fields where large datasets are scarce yet accuracy is critical. In healthcare, Sapient is partnering with leading medical research institutions to deploy HRM to support complex diagnostics, particularly rare-disease cases where data signals are sparse, subtle, and demand deep reasoning. In climate forecasting, HRM raises subseasonal-to-seasonal (S2S) forecasting accuracy to 97 %, a leap that translates directly into social and economic value. In robotics, HRM's low-latency, lightweight architecture serves as an on-device "decision brain," enabling next-generation robots to perceive and act in real time within dynamic environments.

Path Forward

Sapient Intelligence believes that HRM presents a viable alternative to the currently dominant CoT reasoning models. It offers a practical path toward universally capable reasoning systems that rely on architecture, not scale, to push the frontier of AI and, ultimately, close the gap between today's models and true artificial general intelligence.

Availability

The source code is available on GitHub at  https://github.com/sapientinc/HRM.

About Sapient Intelligence

Sapient Intelligence is a global AGI research company headquartered in Singapore, with research centers in San Francisco and Beijing, building next-generation AI models for complex reasoning. Our mission is to reach artificial general intelligence by developing a radically new architecture that integrates reinforcement learning, evolutionary algorithms, and neuroscience research to push beyond the limits of today's LLMs.

In July 2025, we introduced the Sapient Hierarchical Reasoning Model (HRM), a hierarchical, brain-inspired model that achieves deep reasoning with minimal data. With just 27 million parameters and approximately 1,000 training examples, without pre-training, Sapient HRM achieves near-perfect accuracy on Sudoku Extreme, Maze Hard, and other high-difficulty math tasks and outperforms current models that are significantly larger on the ARC-AGI. Early pilot applications will include healthcare, robot control, and climate forecasting.

Our fast-growing team includes alumni of Google DeepMind, DeepSeek, Anthropic, and xAI, alongside researchers from Tsinghua University, Peking University, UC Berkeley, the University of Cambridge, and the University of Alberta, working together to close the gap between today's language models and true general intelligence.

For more information, visit www.sapient.inc.

Media Contact:

genli@sapient.inc
press@sapient.inc

This press release is issued through EmailWire (www.emailwire.com) - the global newswire service that provides effective, local, statewide, national and international press release distribution with guaranteed results.