Reward-predictive items capture attention even when task-irrelevant. While value-driven attention typically generalizes to stimuli sharing critical reward-associated features (e.g., red), recent evidence suggests an alternative generalization mechanism based on feature relationships (e.g., redder). Here, we investigated whether relational coding of reward-associated features operates across different learning contexts by manipulating search mode and target-distractor similarity. Results showed that singleton search training induced value-driven relational attention regardless of target-distractor similarity (Experiments 1a-1b). In contrast, feature search training produced value-driven relational attention only when targets and distractors were dissimilar, but not when they were similar (Experiments 2a-2c). These findings indicate that coarse selection training (singleton search or feature search among dissimilar items) promotes relational coding of reward-associated features, while fine selection (feature search among similar items) engages precise feature coding. The precision of target selection during reward learning thus critically determines value-driven attentional mechanisms.