This study examined two assumptions on the occurrence of on-line inferences: coherence-based and accessibility-based assumptions. To distinguish the predictions of these assumptions, two types of bridging inference conditions were designed. In the case of prediction-consistent condition, contents of bridging inferences matched those of predictions invoked by their prior contexts. In the case of prediction-inconsistent condition, events that were drawn as bridging inference did not match events which were drawn as predictions in their previous contexts. According to the coherence-based assumption, inferences that contribute to establishing coherence of text representations should occur routinely during reading. So, the assumption would expect no differences between prediction-consistent and prediction-inconsistent bridging inferences because both inferences were similarly required for establishing coherence. Following the accessibility-based assumption, the more contextual cues were provided, the more inferences were activated during reading regardless of types of inferences. If this is the case, prediction-consistent bridging inferences would be activated stronger than prediction-inconsistent ones because prediction provides cues to bridging events in prediction-consistent condition but not in prediction-inconsistent condition. As results of priming experiments, accessibility-based assumption was supported. Furthermore, priming effects in prediction-consistent bridging condition were larger than those in sole predictive condition, which suggested additive increment of accessibility for inferences.