LLM Fluid Intelligence

Fluid Intelligence in the context of Large Language Models refers to the capacity for abstract reasoning, problem-solving, and adaptation to novel tasks without relying on pre-existing knowledge or pattern matching of training data. Unlike crystallized intelligence (stored knowledge), fluid intelligence is measured by the ability to generalize from limited examples.

Key Evaluation Benchmarks

ARC-AGI Challenge

The Abstraction and Reasoning Corpus (ARC) serves as a primary benchmark for measuring fluid intelligence. It requires models to solve visual grid-based puzzles that test generalization capabilities rather than memorization.

References