An LLM Builds Poker Abstractions With No Domain Evaluator and Cuts Exploitability 62%

arXiv 2609.04303·medium signalXBlueskyLinkedInCopy link

arXiv 2609.04303 presents Abstraction Agent, a zero-shot pipeline that reads a natural-language game description, discovers continuous strategic features with calibration anchors, scores private states in batches, selects features by correlation and clusters them with k-means, using no hand-strength calculator, training data or game-tree traversal. It reduces lifted-strategy exploitability by up to 62% relative to an expected-hand-strength baseline on heads-up no-limit Texas hold'em turn endgames. It also beat a scalar rank baseline at every granularity on ROVER Trials, an original game absent from any pretraining corpus, which is the ablation that rules out memorization.