1. HRM's Birth and Core Architecture
On August 28, 2025, Live Science reported that Singapore AI company Sapient developed a new AI called the Hierarchical Reasoning Model (HRM), designed to mimic how different brain regions integrate information over millisecond‑to‑minute timescales. Compared with large language models such as ChatGPT that can have tens of billions to trillions of parameters, HRM uses only 27 million parameters and 1,000 training samples, demonstrating higher efficiency and performance. The research paper was uploaded to arXiv on June 26, 2025, and has not yet undergone peer review, but its open‑source code has been published on GitHub for global verification.
At HRM's core is a unique hierarchical architecture composed of two interdependent modules:
- High‑level module: responsible for slow, abstract planning, analogous to strategic decision‑making in the human brain.
- Low‑level module: handles fast, detailed computation, similar to immediate responses in the brain.
Unlike the chain‑of‑thought (CoT) approach used by traditional LLMs, HRM performs sequential reasoning in a single forward pass without explicitly supervising intermediate steps. This 'one‑thought, multi‑step refinement' method improves answer accuracy through iterative refinement: each 'thought' evaluates whether to continue refining or to emit a final answer. This mechanism enables HRM to excel at complex reasoning tasks while reducing computational cost.
2. HRM's Performance Breakthroughs
The Sapient research team validated HRM on the ARC‑AGI benchmark, regarded as a stringent test of progress toward artificial general intelligence (AGI). Results showed:
- ARC‑AGI‑1: HRM scored 40.3%, surpassing OpenAI's o3‑mini‑high (34.5%), Anthropic's Claude 3.7 (21.2%), and Deepseek R1 (15.8%).
- ARC‑AGI‑2: HRM scored 5% on a harder set, ahead of o3‑mini‑high (3%), Deepseek R1 (1.3%), and Claude 3.7 (0.9%).
Additionally, HRM achieved near‑perfect results on complex Sudoku solving and optimal maze pathfinding tasks, where traditional LLMs failed completely (0% accuracy). These achievements derive from HRM's hierarchical design, which overcomes three main CoT shortcomings: fragile task decomposition, high data requirements, and long latency.
3. Innovations That Disrupt Conventional AI
Traditional LLMs like ChatGPT rely on CoT, breaking complex problems into simple natural‑language steps to emulate human thought. The Sapient team noted that CoT suffers from 'fragile task decomposition, enormous data requirements, and high inference latency.' HRM emulates hierarchical brain processing and abandons CoT's text‑based decomposition, reasoning directly in the model's latent space. This 'silent thinking' mode avoids generating intermediate text and significantly improves efficiency.
HRM's training process is also innovative. The researchers report that the main performance gains stem from an 'underreported refinement process' during training rather than the hierarchical architecture alone. The ARC‑AGI benchmark team reproduced the results and confirmed HRM's strong performance, while noting that the architecture's direct effect on performance appears limited. This finding suggests optimizing training methods may be more effective than simply increasing model size.
4. Industry Reaction and Market Potential
HRM's release attracted broad attention in the tech community. X user @AIFanatic called it 'a revolution in reasoning AI,' saying its efficient design could advance AI in resource‑constrained environments such as edge devices and robots. However, some researchers questioned the lack of peer review and warned of potential performance overclaims. Live Science noted that the ARC‑AGI team validated HRM's results using the open‑source code, which enhances credibility.
HRM's potential lies in data efficiency and low compute requirements. Compared with ChatGPT (estimated at 3–5 trillion parameters), HRM's 27 million parameters make it suitable for low‑power devices such as embedded systems and mobile terminals. In fields like medical diagnosis, climate forecasting, and robot control, HRM can handle sparse data to deliver high‑precision reasoning. Sapient plans to apply it to rare disease diagnosis and subseasonal climate prediction and has partnered with medical institutions and climate research teams, indicating broad commercial prospects.
5. Challenges and Future Outlook
Despite strong performance, HRM faces several challenges. First, because reasoning occurs in latent space without CoT's textual intermediate steps, interpretability is limited. The research team demonstrated partial methods for tracing reasoning on some tasks but must further improve transparency to meet regulatory requirements. Second, the lack of peer review may limit academic recognition; Sapient needs to pursue further validation. Also, HRM currently focuses on logical reasoning and lacks the language‑generation capabilities of LLMs; future work should explore hybrid models to cover a wider range of tasks.
Looking ahead, HRM's hierarchical architecture may become a new AI design paradigm. The Sapient team said they are conducting new experiments and expect to report higher ARC‑AGI scores. Combined with reinforcement learning and neuroscience insights, HRM could move toward AGI and challenge the current scale‑dominated AI development path.
6. Conclusion
Sapient's Hierarchical Reasoning Model (HRM), inspired by the human brain, used 27 million parameters and 1,000 training samples to outperform large language models like ChatGPT on reasoning tasks. Its strong results on ARC‑AGI, complex Sudoku, and maze navigation demonstrate the potential of efficient AI. HRM challenges CoT limitations and offers solutions for resource‑limited scenarios. Despite interpretability and validation hurdles, its open‑source release and potential applications make it a focal point in AI for 2025. Industry leaders such as Google and OpenAI should take this emerging architecture seriously, as HRM may reshape the future of AI reasoning.