A position paper suggests the need for a deeper understanding of the algorithms used by Large Language Models (LLMs), as research focus has mainly been on scale and performance improvement.
The proposed AlgEval framework aims to investigate the algorithms LLMs learn and utilize, focusing on algorithmic primitives, attention mechanisms, and inference-time computation.
The framework includes studying the composition of algorithmic primitives to solve specific tasks, with a case study on emergent search algorithms.
The systematic evaluation of LLMs' problem-solving methods can provide insights into internal reasoning and lead to more efficient training methods and novel architectures.