Large language models (LLMs) have shown remarkable ability to tackle complex problems, yet they often stumble on simpler variants of the same tasks. This paradox suggests that LLMs achieve correct outputs through mechanisms fundamentally different from human reasoning. To bridge this gap, researchers have synthesized cognitive science research into a comprehensive taxonomy of 28 cognitive elements. These elements span reasoning invariants, meta-cognitive controls, representations for organizing reasoning and knowledge, and transformation operations.
The study introduces a fine-grained evaluation framework and conducts the first large-scale empirical analysis of 192,000 traces from 18 models across text, vision, and audio. This analysis is complemented by 54 human think-aloud traces, which are made publicly available. The findings reveal that models under-utilize cognitive elements correlated with success, often resorting to rigid sequential processing on ill-structured problems. In contrast, human traces show more abstraction and conceptual processing, while models tend to default to surface-level enumeration.
A meta-analysis of 1,600 LLM reasoning papers indicates that the research community focuses heavily on easily quantifiable elements such as sequential organization (55%) and decomposition (60%), but neglects meta-cognitive controls like self-awareness (16%), which are crucial for success. The study highlights that models possess behavioral repertoires associated with success but fail to deploy them spontaneously.
Leveraging these patterns, the researchers developed test-time reasoning guidance that automatically scaffolds successful structures, improving performance by up to 66.7% on complex problems. By establishing a shared vocabulary between cognitive science and LLM research, the framework enables systematic diagnosis of reasoning failures and principled development of models that reason through robust cognitive mechanisms rather than spurious shortcuts. This work also provides tools to test theories of human cognition at scale.
This research offers a significant step forward in understanding the cognitive foundations of reasoning in LLMs. It provides a roadmap for developing more robust and human-like reasoning capabilities in artificial intelligence, ultimately enhancing the reliability and applicability of these models in real-world scenarios. The insights gained from this study could revolutionize how we approach the design and evaluation of AI systems, making them more aligned with human cognitive processes.
