Assess decoding variability in prompt-based TRP prediction

Systematically analyze how decoding variability, including variation across stochastic generations, affects prompt-based within-turn Transition Relevance Place prediction.

Background

The prompt-based experiments generate a single prediction per interval under a fixed temperature and top-p configuration. This design controls computational cost but does not reveal how robust the predictions are to alternative stochastic generations.

The paper explicitly leaves systematic analysis of decoding variability unresolved. Such an analysis would determine whether the reported prompt-based performance is stable or substantially dependent on the particular sampled generation.