New AI architecture challenges large models on complex reasoning
Research from a Singapore-based AI startup suggests that for complex reasoning, smarter design may be more effective than massive scale. A new research paper on the Hierarchical Reasoning Model (HRM), developed by Singapore-based Sapient Intelligence, shows strong performance on difficult logic problems along with significant improvements in efficiency over current approaches.
The development suggests a promising new direction for AI reasoning, moving away from the "scaling up" trend towards specialised, brain-inspired architectures that prioritise efficiency over scale.
Moving beyond 'thinking out loud'
Most current large language models rely on Chain-of-Thought (CoT) prompting to solve complex problems. This approach forces the model to "think out loud" by generating a text-based, step-by-step monologue before arriving at an answer.
While useful for improving reasoning capabilities, the HRM researchers argue this method has inherent constraints. CoT-based reasoning can be fragile - a misstep in the generated text can derail the entire process. It's also computationally intensive, as the model's reasoning becomes tied to the serial, token-by-token generation of language.
"CoT for reasoning is a crutch, not a satisfactory solution," the researchers write. "It relies on brittle, human-defined decompositions where a single misstep can derail the reasoning process entirely."
A brain-inspired approach to internal reasoning
HRM takes a fundamentally different path, inspired by how the human brain organises computation across different timescales. Instead of thinking in text, it performs "latent reasoning" - solving problems in its internal, abstract representation without translating each step into language.
The architecture uses two coupled modules operating at different speeds: a high-level 'planner' module that operates slowly to set overall strategy, and a low-level 'worker' module that performs rapid, detailed calculations to execute the current part of the plan.
This structure enables what the team calls "hierarchical convergence." The fast worker module addresses portions of the problem until it reaches a stable solution, then the slow planner takes this result, updates its strategy, and gives the worker a new, refined sub-problem. This process allows the model to perform long sequences of reasoning steps without the computational pitfalls that plague other approaches.
Breakthrough results on intractable problems
The model was tested on benchmarks requiring extensive search and backtracking, including extremely difficult Sudoku puzzles, complex mazes, and the Abstraction and Reasoning Corpus (ARC-AGI) - a key test of abstract reasoning ability.
The results are striking. On tasks where leading CoT models scored 0% accuracy, HRM achieved near-perfect performance after being trained on just 1,000 examples. On the ARC-AGI benchmark, the 27-million-parameter HRM scored 40.3%, outperforming much larger models including OpenAI's o3-mini-high (34.5%) and Claude 3.7 Sonnet (21.2%).
According to Sapient Intelligence CEO Guan Wang, this efficiency could translate to significant speedup in task completion time compared to CoT methods - potentially up to 100x faster for the specific reasoning tasks where HRM excels.
And there are gains in training, where professional-level Sudoku training required approximately two GPU hours, a fraction of the resources needed for foundation models.
Implications for research and enterprise
This level of performance from a small, efficient model has significant implications for university research and operations. Wang suggests that while large language models remain optimal for creative and language-based work, HRM-like architectures are superior for "complex or deterministic tasks" requiring sequential decision-making or long-term planning.
This includes applications in robotics, scientific exploration, and optimising complex systems - areas directly relevant to research and enterprise functions. The model's ability to run on standard hardware with minimal memory requirements opens possibilities for real-time reasoning in data-scarce environments where computational budgets are limited.
Looking ahead
Sapient Intelligence is already working to evolve HRM from a specialised problem-solver into a more general-purpose reasoning module, with promising initial results in healthcare, climate forecasting, and robotics applications.