Are LLMs as smart as humans?

Since LLMs are trained on a large amount of data, they incorporate knowledge about the world in many domains and often demonstrate surprising capabilities by “demonstrating more general intelligence than previous models.” [22] Generative AI models demonstrate close to human performance on many tasks, including passing exams, even difficult ones such as law or medical exams [23, 24], or passing mock interviews on Leetcode [22]. But passing an exam doesn’t mean that generative AI is ready to act as a lawyer or a doctor.

To answer this question, there have been multiple research projects on LLM reasoning capabilities. Researchers typically attack this problem head-on. One idea is to develop novel and complex prompting techniques that enhance the reasoning capabilities of such models – such as CoT, self-consistency with CoT, graph-of-thought, tree-of-thought, and others. Another is to evaluate LLM capabilities on various tasks...