Learning Science
Feedback: The Highest-Leverage Variable in Retrieval Practice
Retrieval practice without feedback teaches almost nothing. What feedback contains, and when it is delivered, changes the outcome more than almost any other variable in this literature.
No feedback, low success: worth nothing
As shown on the testing-effect page, Rowland (2014, Psychological Bulletin, 159 effect sizes) found that a retrieval attempt with initial success below 50% and no feedback produces g = 0.03, a confidence interval of [−0.21, 0.27] that includes zero. A failed retrieval attempt with no correction is not a smaller version of the testing effect — it is indistinguishable from no effect at all.
What kind of feedback works
Van der Kleij, Feskens & Eggen (2015, Review of Educational Research, 40 studies, 70 effect sizes) compared three feedback formats directly:
Elaborated feedback's advantage was disproportionately larger for higher-order outcomes than for simple factual recall. This is corroborated by Yang, Luo, Vadillo, Yu & Shanks (2021): corrective feedback produced g = 0.537 versus g = 0.374with no feedback. It is also corroborated by Pan & Rickard (2018), where elaborated retrieval is one of only three moderators that survive both simultaneous fitting and publication-bias correction (Δd = 0.22–0.23).
Overall effect and its limits
Wisniewski, Zierer & Hattie (2020, Frontiers in Psychology10:3087; 435 studies, 994 effect sizes, N > 61,000) found an overall feedback effect of d = 0.48, with impact substantially moderated by how much information the feedback actually contained. Feedback was measurably weaker on motivational and behavioral outcomes than on cognitive ones — feedback that changes what someone knows is more reliable than feedback intended to change how they feel about learning.
Delayed feedback: the intuitive claim doesn't hold up
Not supported
A commonly repeated claim holds that delaying feedback (rather than giving it immediately) improves long-term retention — the "delayed-retention effect." The classic version of this finding is confounded by uncontrolled lag-to-test, and the adult effect disappears once that lag is controlled (Metcalfe, Kornell & Finn 2009). Van der Kleij, Feskens & Eggen (2015) found delay to be a negative moderator in their meta-analysis. Ryan et al. (2024, Medical Education) ran a randomized controlled trial directly testing immediate versus delayed feedback and found no discernible difference, recommending that effort go toward feedback quality instead of feedback timing.
Hypercorrection: confidence changes how feedback lands
As described on the self-assessment page, Butterfield & Metcalfe (2001, Journal of Experimental Psychology: Learning, Memory, and Cognition, mean within-subject G = .36) show that learners who were highly confident in a wrong answer absorb the correction to that answer especially efficiently. Metcalfe & Finn (2011) found high-confidence errors preferentially resurface at a one-week delay if uncorrected. The same mechanism that makes hypercorrection useful for fixing confident errors means confidently wrong feedback is absorbed just as efficiently — a corrected wrong answer is not a neutral non-event.
A caution for AI-graded feedback specifically
Documented failure modes
Zheng et al. (NeurIPS 2023) found GPT-4 acting as a grading judge exceeded 80% agreement with human graders — matching typical human-to-human agreement rates. The same paper documents four specific failure modes: position bias, verbosity bias (a longer answer receives more credit independent of correctness), self-enhancement bias, and — notably — the judge's weakest performance is specifically on reasoning and math grading. Verbosity bias in particular is a direct, measurable corruption of any system that uses graded performance as a record: a longer low-quality answer can outscore a shorter correct one unless grading is anchored to an explicit, enumerated rubric rather than holistic judgment.