PFIA 2024
JIAF
Exploring Emergent Skills with Chess-GPT
We explore game strategies in Chess within Large Language Models (LLMs) using games in Portable Game Notation (PGN) from the Lichess database as training data. Our objective is to examine the capacity of LLMs to develop new Skills, which may be considered as a class of emergent properties. We investigate the ability to solve Chess puzzles (games that require a unique correct move) and seek to assess the success rate of an LLM in performing this task. Subsequently, we study how this success rate changes when we alter the elo parameter context in the header of the PGN.