Agent-based models explain how opinions evolve through interaction, yet they often remain detached from real data. This work develops a probabilistic inference framework that learns latent opinions and model parameters directly from social traces, retaining the interpretability of theoretical models while gaining empirical grounding. Applied to political conversations on Reddit, it provides a data-driven test of opinion dynamics theories, revealing limited evidence for the backfire effect in online debate.