Cognitive Foundations for Reasoning and Their Manifestation in LLMs
3 days ago
- #Large Language Models
- #Cognitive Science
- #Artificial Intelligence
- Large language models (LLMs) solve complex problems but fail on simpler variants, indicating different reasoning mechanisms from humans.
- A taxonomy of 28 cognitive elements is synthesized from cognitive science research to analyze reasoning behaviors in LLMs.
- A fine-grained cognitive evaluation framework is proposed, analyzing 170K traces from 17 models and 54 human think-aloud traces.
- Systematic structural differences are found: humans use hierarchical nesting and meta-cognitive monitoring, while models rely on shallow forward chaining.
- Meta-analysis of 1,598 LLM reasoning papers shows research focuses on easily quantifiable behaviors, neglecting meta-cognitive controls correlated with success.
- Models possess behavioral repertoires associated with success but fail to deploy them spontaneously.
- Test-time reasoning guidance is developed to scaffold successful structures, improving performance by up to 60% on complex problems.
- The study bridges cognitive science and LLM research, aiming for models that reason through principled cognitive mechanisms rather than shortcuts or memorization.