Fluid Intelligence
Fluid intelligence is a concept from psychologist Raymond Cattell's two-factor theory of intelligence, referring to the capacity for abstract reasoning, pattern recognition, and novel problem-solving without reliance on prior experience. It stands in contrast to crystallized intelligence, which depends on accumulated knowledge. In AI, it is used to measure a model's ability to handle problems outside its training distribution.
Timeline (last 90 days)
Fluid intelligence refers to the ability to reason, discover patterns, and adapt to change without relying on prior experience, as opposed to crystallized intelligence, which depends on accumulated knowledge.
The distinction between fluid intelligence and crystallized intelligence comes from psychologist Raymond Cattell's two-factor theory of intelligence, proposed in the 1940s.
The original argument claims that the fluid intelligence level of base LLMs to date remains close to zero, whereas LRMs possess a considerably substantial degree of fluid intelligence.
Simply stacking compute and scaling parameters cannot give transductive models fluid intelligence; it is the paradigm shift from guessing answers to constructing reasoning programs that brings about a qualitative leap.
Traditional knowledge-based benchmarks may mask the absence of fluid intelligence, while tasks such as ARC are better at exposing paradigm differences.
All Facts (5)
Fluid intelligence refers to the ability to reason, discover patterns, and adapt to change without relying on prior experience, as opposed to crystallized intelligence, which depends on accumulated knowledge.
50%UnverifiedThe distinction between fluid intelligence and crystallized intelligence comes from psychologist Raymond Cattell's two-factor theory of intelligence, proposed in the 1940s.
50%UnverifiedThe original argument claims that the fluid intelligence level of base LLMs to date remains close to zero, whereas LRMs possess a considerably substantial degree of fluid intelligence.
50%UnverifiedSimply stacking compute and scaling parameters cannot give transductive models fluid intelligence; it is the paradigm shift from guessing answers to constructing reasoning programs that brings about a qualitative leap.
50%UnverifiedTraditional knowledge-based benchmarks may mask the absence of fluid intelligence, while tasks such as ARC are better at exposing paradigm differences.
50%