Context-dependent reinforcement learning (RL) challenges the assumption that decision makers encode the absolute values of choice outcomes. This study investigates whether the associated choice biases arise from a relative encoding of outcomes or an alternative mechanism involving cumulative reward learning and selective attention to outcomes. Using eye tracking, participants completed a RL task where choice options were initially learned in fixed contexts before being tested in novel pairings. Results revealed an overall preference for options that were contextually favored in the learning phase, even when these preferences violated expected value maximization. Computational model comparisons demonstrated that hybrid encoding models, incorporating absolute and relative values, provided the best overall account of individual behavior. While eye fixations on choice outcomes decreased over trials, fixation-dependent RL models did not fit the data well, suggesting that overt visual attention patterns do not fully explain context-dependent choice biases.