Pain expectations are formed by learning the association between specific circumstances and the occurrence of pain. These expectations are known to influence pain perception. Whether this learning is statistically optimal remains debated. If it is, the learning process of the participant should reflect their perception of the precision of each observation (stochasticity) and of the instability of the environment (volatility).Further, it is unclear whether inter-individual differences observed in experimental pain learning studies are due to temporary states or intrinsic traits.To probe these questions, we administered a novel pain reversal learning task (with periods of high and low volatility) twice (1 week apart) to 50 healthy volunteers. The task, in which participants had to learn the probabilistic (0.2/0.8) association between 2 arbitrary visual cues and two heat stimulus intensities (pain/no pain), included 160 trials and 10 reversals. Computational modeling of the participants’ responses and Bayesian model selection are used to test specific hypotheses on pain learning strategies. ICC coefficients will be used to quantify the test-retest reliability of the best model’s learning parameters.Data collection is complete but modeling is still ongoing.Preliminary results, derived from a reduced set of models, suggest that perceived volatility influences human pain learning (p=0.0005), suggesting quasi statistically optimal learning. Conversely, ICC coefficients indicated poor-to-excellent test-retest reliability of pain learning parameters (ICC 95% CIs ranging from <0.4 to 1), meaning that some pain learning parameters might reflect intrinsic traits whereas others would be related to transient states.