Research

18 October 2026

Does Outcome-Value Supervision Add Decision-Relevant Information to a Compact Chess Policy?

A compact chess network proposes four moves and its value head reranks them, without tree search. In a prespecified two-seed ablation, training with a value-loss weight of 0.5 instead of 0 made a Stockfish evaluation easier to decode from the network's trunk and improved reranking once the value head was retrained. Top-1 accuracy of the policy itself on puzzles did not improve consistently.

• chess
• policy-value networks
• reranking
• ablation
• representation probing