Hierarchical Reasoning Models Enhanced by Recurrent Networks: A New Approach for Logical Thinking
In the realm of artificial intelligence (AI), a new breed of models is making waves with its ability to reason effectively and solve complex problems in a manner reminiscent of human cognition. These models are known as Hierarchical Mind Models (HRMs).
The development of HRMs can be traced back to the desire to understand complex cognitive processes by structuring mental functions in layered levels, mirroring how the human brain organises information from basic to higher-order reasoning. The HRM workflow involves a trainable embedding layer, transforming the input into a machine-readable form.
At its core, an HRM consists of two coupled recurrent modules: a high-level module for abstract reasoning and a low-level module for quick, complex calculations. The low-level module takes several quick steps to reach a partial solution, updates its hidden state, and sends the result up to the high-level module.
The high-level module, in turn, demonstrates effective reasoning through structured recurrence inspired by the human brain. It updates its plan based on the low-level module's result and sends a new high-level hidden state back down to the low-level module. This cyclical process continues until both modules agree on the final answer, with the low-level module running for NxT times, where N is the number of times the high-level module runs.
One of the key advantages of HRMs is their ability to exhibit exceptional problem-solving skills in logical and navigation tasks, without relying on extensive pretraining like foundation models. They also avoid issues with memory usage and training stability in recurrent neural networks by using a one-step gradient approximation.
Moreover, HRMs combine high-level planning with quick low-level computation, outperforming some of the most sophisticated Language Models (LLMs) available today. The high-level module's final hidden state does not go directly into the output layer; instead, it goes through a halting head that determines whether the model should stop or continue for another N cycle.
As we continue to explore and refine HRMs, they may very well define the next phase of AGI research, bringing us one step closer to AI that thinks more like humans. With their potential to revolutionise the field of AI, the future of HRMs is undoubtedly an exciting one.
