Misconceptions in LLMs
Subbarao highlights the flawed assumptions surrounding the capabilities of LLMs, particularly the misconception that they excel at verification due to the perceived simplicity of the task compared to generation. He argues that LLMs operate in constant time for generating responses, regardless of the complexity of the underlying problem, and critiques the idea of introducing pause tokens as a means to encourage reasoning.In this clip
From this podcast
Machine Learning Street Talk (MLST)
Prof. Subbarao Kambhampati - LLMs don't reason, they memorize (ICML2024 2/13)
Related Questions