What does a chess model actually evaluate?
Question
In your words: what does the model actually evaluate?
Method
In your words: square ablation. Remove squares and measure how the evaluation changes
Chart placeholderchange in the model's evaluation when each square is removed (heatmap over the board)
Finding
In your words: the model put high probability on a weak move that looks natural
Chart placeholdermodel probability vs. engine evaluation for the candidate moves
Limitation
In your words: what single-square ablation can't see
So what
In your words: why plausible isn't the same as correct
- Code: link needed
- Notebook: link needed