reasoningfactbullishReinforcement learning enabled chess models to eliminate checkmate confabulation entirely and increase legality from 48% to 61%Computation and Language28 Jul 2026http://arxiv.org/abs/2607.20952v1