Can an LLM add anything beyond chess statistics?

Same chess position

One real decision from a held-out game

Statistical baseline

Engine / statistical evidence

Maia-3

Human move plausibility

Qwen3-8B

LLM move preference

Compare + combine

Same positions, same held-out evaluation

Does Qwen add signal?

Maia-3 is not an LLM. It is a chess-specific, rating-conditioned human-move transformer, used here as a human-behavior control: it tells us how much of Qwen's preference is simply “what humans play”. The question is not which model is best; it is whether Qwen carries information the other two do not.

← → keys also work