People’s posting choices on online social media platforms shape not only what others see, but also what spreads within these social networks, including ideas, beliefs, and behaviours. Previous research has focused on what kinds of content tend to be shared, but less is known about how individuals, through reward (such as likes) and social influence, learn what to post over time. In this project, we will model posting behaviour as a reward and social learning process using the social feature learning (SFL) model. The SFL model describes how individuals learn to weigh social and non-social features of their environment to maximise their own rewards, providing a dynamic account of human social learning. While previous studies tested SFL in controlled experiments, here we will test whether it can also explain naturalistic human behaviour in online social networks. To ensure the feasibility of our approach, we have analysed an initial subset of our data, containing 3.3 million posts from 978 Bluesky users. Each post was represented as a probability distribution over 100 topics, identified using transformer-based topic modelling. To test different social learning strategies, we extracted social features from each user’s feed, capturing average popularity and engagement of topics, along with the popularity of the posting users. In our preliminary analysis, we compared two models of posting behaviour, a basic asocial reinforcement learning (RL) model and the SFL model. Across users, we find that the SFL model provides a better account of posting behaviour (mean BIC weight = 0.83), indicating that posting decisions are shaped not only by rewards but also by various social learning strategies, and that SFL can capture aspects of human social learning in the wild. We are currently extending these analyses to the full dataset (~4,000 users, >8 million posts) and will systematically test additional model variants and combinations. We will then use the calibrated models as the basis for agent-based simulations, embedding SFL-controlled agents in large simulated networks. This combination of large-scale model calibration and simulation will link individual posting decisions to emergent network dynamics in a more realistic and empirically grounded way.