This paper presents two studies comparing the effects of educational immersive virtual reality (IVR) versus traditional videos on conceptual knowledge and self-efficacy. Learning was measured through multiple-choice questions assessing conceptual knowledge, and open-ended questions assessing knowledge understanding, knowledge application, and knowledge creation, based on Bloom’s Taxonomy of learning objectives. In Experiment 1, 53 eighth grade students from an all-girls school learned about humans’ impact on the ocean through either 360-degree videos, using a virtual reality headset, or through traditional videos, using a computer monitor. Measures were taken before and right after treatment. In Experiment 2, 139 sixth-to-eighth grade students from the same school used the same instructional material and equipment from Experiment 1, and measures were taken four times (before treatment, right after two treatment sessions, and five weeks after treatment). Also, we measured learning agency and investigated its mediation role between condition and self-efficacy. The groups did not differ on the multiple-choice scores assessing conceptual knowledge in Experiment 1. In Experiment 2, participants in the IVR group scored higher for knowledge creation than participants in the desktop group, but not knowledge understanding or knowledge application. The IVR group scored higher on self-efficacy than the desktop group in Experiment 1, but not in Experiment 2. Finally, learning agency mediated the relationship between condition and self-efficacy in Experiment 2, indicating a possible mechanism underlying immersion effects on self-efficacy. Results are discussed in light of cognitive sciences and their implication for learning in immersive virtual reality.