llm-kb
← Back to research
Paper

AlphaZero Auxiliary Loss improves play in sparsely rewarded games

A new paper studies the gap between strong and perfect play in AlphaZero for Connect Four and Chomp, introducing AlphaZero Auxiliary Loss (AZAL) that adds oracle-derived policy supervision. The work highlights limits of vanilla AlphaZero in sparsely rewarded games and shows auxiliary supervision can narrow the gap.

0 engagement·1 source·Thu, Jul 9, 2026, 11:17 PM
The paper, posted on arXiv on 2026-07-09, evaluates AlphaZero in two oracle-evaluable domains: Connect Four (solved partisan game) and Chomp (impartial game with Grundy-number structure). Under a unified self-play + MCTS pipeline, they compare vanilla AlphaZero, a multi-frame variant (limited to Chomp), and AZAL. AZAL incorporates oracle-derived policy supervision to improve learning. The study provides concrete evidence that strong play does not imply perfect play and that auxiliary losses can help.

Entities

AlphaZero(model)AlphaZero Auxiliary Loss (AZAL)(concept)Connect Four(concept)Chomp(concept)Monte Carlo Tree Search(concept)

Related