Drift-diffusion models (DDMs), which are a type of process model, have shown promising potential in predicting the choices of consumer products, risky prospects, and intertemporal alternatives in recent years. The basic assumption underlying all DDMs is that preferences are constructed through a dynamic information acquisition process. Specifically, individuals’ preferences are formed by sampling from available alternatives until the evidence supporting one alternative is strong enough to induce a decision. Two key debates exist among different versions of DDMs. One is the issue of the causal link between gaze and decision. Certain DDMs assume that visual fixation process has a causal effect on choices, whereas other DDMs assume that fixation lengths are affected by the accumulated evidence of the alternatives. Such evidence produces a correlation between fixation lengths and choices without giving a causal role to fixations. The other key debate is the primacy/recency issue. Hence, incoming evidence is weighted more (recency) or less (primacy) than previously accumulated evidence. These two issues are unresolved and require further evidence. In the present study, we employed a gaze-contingent paradigm to manipulate participants’ gaze in a value-based binary food choice context. This paradigm involves endogenously manipulating the timing of decision s. By monitoring participants’ eye movements during decision-making, an experimenter can terminate participants’ deliberation and them to choose between two options the moment they fixate on a specific target option for a set amount of time. First, this gaze-contingent manipulation allowed us to examine the causal link between gaze patterns and choices. Therefore, we examined if participants’ choices were systematically biased toward the target options. Second, the gaze-contingent manipulation has different influences on the gaze time in various decision periods. Hence, we also investigated the dynamic changes of the decision evidence weighting to examine the primacy/recency assumption of DDMs. In the current study, we found that, first, participants’ value-based choices were systematically biased to the predetermined target options. Second, the relative time advantage and the last fixated option could significantly predict participants’ choices. Third, the gaze-contingent paradigm affected the gaze duration in the last third of the decision process. Fourth, causal mediation analysis revealed that the relative time advantage in the last third period could fully mediate the effect of target option on choices, i.e., the predetermined target options lead to longer duration on the target options in the last third period, thus making participants likely to choose the target options. Fifth, the results of the estimated parameters supported the recency assumption of DDMs, indicating that incoming evidence is weighted more than previously accumulated evidence. Results generally suggest that a causal link exists between gaze-contingent manipulation and value-based choices and that recent evidence is weighted more during the evidence accumulation process in DDMs. Moreover, findings revealed the mechanism of the gaze-contingent manipulation paradigm and the dynamic fixation process during this paradigm. This research helps clarify the two key issues of DDMs and deepens our understanding of the relationship between visual fixation and value-based choices.