Position: System-2 AI is about Complexity Out of Distribution

Mahdi Samiei, Mohamadreza Fereydooni, Yousef Javaherian, Mahdieh Soleymani Baghshah

Abstract: This position paper argues that addressing the limitations of the current "System-1" paradigm in deep learning, which struggles to generalize to complex scenarios beyond training, necessitates the introduction of a complementary "System-2" reasoning paradigm. We introduce the concept of "complexity out-of-distribution," which highlights the obstacles in progressing toward true artificial general intelligence (AGI). These scenarios require more intricate representations or computational paths than those encountered during training. Our position is that achieving effective solutions for such out-of-distribution complexities calls for a shift towards System-2, which frames problem-solving as a search over sequences of semantic units with unbounded complexity. This new paradigm seeks to discover algorithms, leveraging System-1’s learned representations and heuristics, to handle examples with varying complexity akin to human reasoning abilities. We assert that advancements necessitate the development of tailored System-2 methods, including complexity-focused tasks, benchmarks, supervision paradigms, representations, metrics, and inductive biases. By drawing on recent research across multiple domains, we outline the essential requirements and challenges in integrating the symbolic search process of System-2 with neural network architectures.

Primary Area: Research Priorities, Methodology, and Evaluation

Keywords: system-2, out of distribution generalization, reasoning

Submission Number: 473