Public articles linked to the same research event.
arXiv Using controlled two-player Rock–Paper–Scissors games (statistical versus Markov players, 100 to 1000 rounds of context) and a one-player stochastic n-gram continuation task (orders 1 to 8), the authors separate strategy identification, marginal distribution matching, and conditional rule execution, finding that Markov strategies are harder to identify than non-Markov ones, longer context does not improve and often degrades identification, correct recognition does not guarantee faithful simulation, and second-order dependencies markedly lower strict rule match, while neither providing transition rules nor longer prefixes removes high-order degradation and a prefix-only empirical n-gram estimator stays comparatively stable at high order.
Using controlled two-player Rock–Paper–Scissors games (statistical versus Markov players, 100 to 1000 rounds of context) and a one-player stochastic n-gram continuation task (orders 1 to 8), the authors separate strategy identification, marginal distribution matching, and conditional rule execution, finding that Markov strategies are harder to identify than non-Markov ones, longer context does not improve and often degrades identification, correct recognition does not guarantee faithful simulation, and second-order dependencies markedly lower strict rule match, while neither providing transition rules nor longer prefixes removes high-order degradation and a prefix-only empirical n-gram estimator stays comparatively stable at high order.
Using controlled two-player Rock–Paper–Scissors games (statistical versus Markov players, 100 to 1000 rounds of context) and a one-player stochastic n-gram continuation task (orders 1 to 8), the authors separate strategy identification, marginal distribution matching, and conditional rule execution, finding that Markov strategies are harder to identify than non-Markov ones, longer context does not improve and often degrades identification, correct recognition does not guarantee faithful simulation, and second-order dependencies markedly lower strict rule match, while neither providing transition rules nor longer prefixes removes high-order degradation and a prefix-only empirical n-gram estimator stays comparatively stable at high order.
Using controlled two-player Rock–Paper–Scissors games (statistical versus Markov players, 100 to 1000 rounds of context) and a one-player stochastic n-gram continuation task (orders 1 to 8), the authors separate strategy identification, marginal distribution matching, and conditional rule execution, finding that Markov strategies are harder to identify than non-Markov ones, longer context does not improve and often degrades identification, correct recognition does not guarantee faithful simulation, and second-order dependencies markedly lower strict rule match, while neither providing transition rules nor longer prefixes removes high-order degradation and a prefix-only empirical n-gram estimator stays comparatively stable at high order.