Some Assembly Required: Self-Testing Alone is Not an Effective Way to Build Broader Understanding

Extensive research has demonstrated that testing enhances long-term memory by strengthening retrieval processes for the tested memory traces. Such memory models explain why learners benefit from repeated testing but do not account for how learners construct understanding across multiple ideas. In contrast, theories of discourse comprehension emphasize the construction of inferences, whereby learners actively connect ideas into a coherent mental representation. These two perspectives—memory-based retrieval versus inference-based integration—make distinct predictions about what learners retain and understand. The present study tested these predictions by comparing self-testing, self-explanation, and reading strategies on both fact memory and inference performance. In Experiment 1 (n = 254), despite superior retention of individual facts, self-testing produced lower inference performance than self-explanation. Experiment 2 (n = 68), a preregistered replication with a minimum 24-hour delay, confirmed and extended these findings: repeated retrieval improved recall of facts but did not facilitate the assembly of coherent knowledge structures. A mixed-methods analysis revealed that generating inferences during learning specifically predicted inference performance, whereas noting facts predicted fact recall, providing converging evidence that these strategies produce qualitatively different mental representations. Furthermore, participants generated fewer inferences when self-explaining science texts compared with narratives, highlighting domain differences in the ease of knowledge integration. Together, these results demonstrate that while self-testing strengthens memory for discrete information, it fails to foster the integrative processing necessary for comprehension. Effective learning requires more than strengthening retrieval processes—it depends on the active generation of inferences that bind ideas into coherent mental models.