Evaluating the Intelligence of large language models: A comparative study using verbal and visual IQ tests.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Machine Learning Evaluation

Background:

• Large language models (LLMs) demonstrate proficiency on specialized tasks, but their general reasoning capabilities require further investigation.
• Existing benchmarks may not fully capture the breadth of cognitive abilities in advanced AI systems.

Purpose of the Study:

• To evaluate the general reasoning abilities of 18 diverse large language models using a comprehensive IQ test suite.
• To analyze the impact of model scale, multimodality, and multi-agent reflection on reasoning performance.
• To establish IQ tests as a standardized benchmark for longitudinal comparison of LLM cognitive abilities against human norms.

Main Methods:

• Administered a 14-section IQ suite covering verbal, numerical, and visual reasoning tasks to 18 LLMs.
• Implemented a multi-agent reflection variant where models critique and revise answers.
• Analyzed performance variations based on model size, multimodality, and reasoning task type.

Main Results:

• Observed a significant bias towards verbal reasoning (e.g., GPT-4: 79% accuracy) over numerical reasoning (53% accuracy).
• Identified a pronounced modality gap, with text-based IQ scores (≈125) significantly higher than visual-based scores (≈103).
• Found persistent difficulties in abstract arithmetic tasks (≤20% accuracy) and modest gains from multi-agent reflection in frontier models.

Conclusions:

• LLM intelligence, as measured by IQ tests, scales with model size but exhibits non-uniform gains across different reasoning domains.
• IQ tests provide a valuable, human-referenced framework for evaluating and comparing LLM cognitive abilities, despite limitations.
• Further research is needed to develop more comprehensive AI evaluation benchmarks that capture nuanced reasoning capabilities.