Open Internet by MindsNet

Understanding Model Response to Impossible Chess Moves

The challenge is to determine how a chess-playing language model responds to impossible moves, and whether its internal board state representation is causal or epiphenomenal. This involves analyzing the model's behavior under different types of impossible moves, such as rule violations, trajectory violations, impossible threats, and referential ambiguity.

Computing & Technology, Computer Science, Machine Learning