CS-RNR lets AI agents certify their own exploits in games

Researchers Boning Li and Longbo Huang have published a new approach to opponent exploitation in imperfect-information games, called budget-constrained confidence-scheduled restricted responses (CS-RNR). Traditional Nash-equilibrium strategies guarantee the game value but ignore extra profits available by exploiting flawed opponents. However, aggressive best responses to an incomplete opponent model can backfire. CS-RNR instead tracks opponents' pooled action frequencies using anytime-valid confidence sequences, flagging actions as exploitable only when their intervals separate from an equilibrium reference. These confirmed deviations form a conservative opponent model, which is solved into candidate counter-strategies across a grid of pin levels.

Before deployment, every candidate is evaluated by a full-tree best response, producing a certificate that is compared against a user-specified budget and committed atomically with the strategy. This makes safety a property of the actual played strategy, not just the training method. In Leduc hold'em, CS-RNR delivers 6.2x the steady-state gain of a money-verified binary gate while respecting budget constraints. A trajectory mixture using the same estimator reaches 13.6x the budget. Across Leduc, Liar's Dice, and 5-rank Leduc, all 36,000 audited hands satisfy the reported certificate tolerance, demonstrating strong empirical safety with meaningful exploitation gains.