Training variability facilitates the flexibility of goal-specific spatial probability learning.

Selection history is traditionally characterized as a rigid and automated habit that persists regardless of current goals. Although recent research suggests that explicit goals can modulate this process, the mechanisms underlying this flexibility and its acquisition conditions remain poorly understood. Across two experiments using simple geometric shapes, the present study examined the joint modulatory effects of explicit goals and training variability on spatial probability learning. Experiment 1 demonstrated that explicit goals can directly activate the spatial priority map, ruling out confounds based on long-term semantic associations. Crucially, this modulation was found to be highly target-specific, as the spatial bias vanished immediately when the explicit cue was removed. Experiment 2 investigated the acquisition of this flexibility through the lens of the Contextual Interference effect. We found that training format played a highly influential role: In Experiment 2 A (Interleaved training), where background colors switched randomly across trials, participants successfully established a robust, flexible "context-goal" mapping. However, in Experiment 2B (Blocked training), participants failed to transfer the learned bias to a dynamic test phase, despite showing strong effects during learning. These results suggest that flexibility is not inherent to probability cueing but is learned. Establishing a robust Contextual Gating mechanism requires high-variability training that forces active strategy retrieval, whereas low-variability training yields only fragile biases dependent on passive priming. These findings support a conditional automaticity account of selection history.