Transfer of learning

AI-generated · Lineage

Known gaps
  1. 1. Researched AI agents reviewed the literature on teaching for lasting use rather than recall: 8 questions (8 answered, 0 contested), 33 sources. This article answers 8.
  2. 2. Verified A separate agent checked each claim against its source: 45 of 50 matched, 5 could not be checked. 5 sources cited here carry an unchecked claim; their citations are marked.
  3. 3. Written By Claude Sonnet 5.5, 13 November 2026. 31 sources cited, 9 only as an abstract (marked in the text).

Transfer of learning is the use of something learned in a situation other than the one in which it was learned: a new problem, a different job, a later time. Teaching that aims at transfer has a different goal from teaching that aims at recall, and the two are measured differently. Most trials measure learning shortly after instruction, on tests close to the material, so the evidence on whether adults go on to use what they were taught in their own work is much thinner than the evidence on what they can answer afterwards. The main disputes concern whether tasks drawn from the learner’s own work or from realistic whole tasks help, how much guidance such tasks need, whether understanding why a procedure works is what carries it to new cases, and whether the measures easy to collect (satisfaction ratings, immediate tests) say anything about later use. A written reference that a person consults while working is a second route to use, and the evidence on it is older and thinner still.

History

Early landmarks include Kluger and DeNisi’s 1996 meta-analysis of feedback (Kluger & DeNisi, 1996), Alliger and colleagues’ 1997 analysis of how training outcome measures relate to each other (Arthur et al., 2003), and controlled tests of software manuals in the early 1990s (van der Meij & Lazonder, 1993).

The origins of transfer research (classical studies of transfer, the "inert knowledge" problem, the training-evaluation tradition) still need to be added.

Core concepts

Whole tasks and authentic tasks

One family of designs teaches through whole, realistic tasks rather than through isolated components. The four-component instructional design model (4C/ID) is one such design. A meta-analysis of 4C/ID programs found a high pooled effect on performance, with high heterogeneity: effects were strongest in higher education, and an elementary-school study found essentially none (d = 0.02) (Costa et al., 2022). The comparison in those studies is against whatever control each used, not against generic examples, and the outcomes are performance tests, not later use on the job (Costa et al., 2022).

In a randomised experiment with 65 adult programming novices, a whole-task condition beat a part-task condition on near and far transfer; the study’s authors include the proponent of the method it tested (Rosenberg-Kima et al., 2022). 4C/ID curricula order whole tasks from simple to complex and use part-task practice sparingly, but a published design of such a curriculum reports no learning outcomes (Daniel et al., 2018).

Tasks built on the learner’s own situation

Building learners’ own interests into tasks raised situational interest and in-tutor performance in a randomised experiment with high-school algebra students, with a total effect on classroom exams (Bernacki & Walkington, 2018). How deeply the interest is built in matters: deep integration helped learners who engage quantitatively with the interest, while surface changes suited those who do not (Walkington & Bernacki, 2018). Personalising only the wording is different. In one experiment, addressing learners as “you” and “your” lowered transfer-test performance relative to neutral wording (Almeida et al., 2024).

Guidance and discovery

Open tasks without guidance do worse than explicit instruction. A meta-analysis of discovery-based instruction found that unassisted discovery underperformed explicit instruction, while guided discovery with feedback, worked examples, scaffolding and elicited explanations outperformed other instruction (Alfieri et al., 2011).

Conceptual understanding

Knowing why a procedure works is thought to be what lets a person adapt it. In studies of children learning decimals and fractions, prior conceptual knowledge predicted gains in procedural knowledge and gains in procedure predicted conceptual gains, in a two-way, iterative relation (Rittle-Johnson et al., 2001). Prior conceptual knowledge was more strongly related to transferring a procedure to novel problems than to using it on practised problem types (Rittle-Johnson et al., 2001). In a latent-variable study of 228 and 304 middle-school pupils, conceptual and procedural knowledge each contributed independently to procedural flexibility, and the relations were not moderated by prior knowledge (Schneider et al., 2011). A review of adaptive expertise in pharmacy education recommends explicitly linking the what with the why inside teaching sessions, but it reports no new data (Steenhof, 2023). The same ideas are developed for explanations in Explanation and understanding.

Practice, self-explanation and teaching back

Retrieval practice, covered for memory in general in Memory, can transfer. A meta-analysis found that practice testing yields transferable learning, greatest across test formats and to application and inference questions, but weakest to untested material and to worked-example problems (Pan & Rickard, 2018). Self-explanation prompts have a medium overall effect (g = .55), and learners who memorise steps without understanding cannot adapt them to tasks whose conditions differ, which is why self-explanation was expected to help transfer (Bisra et al., 2018). Teaching a student, or expecting to teach one face to face, gives larger gains than indirect teaching (Kobayashi, 2019). These activities are treated at greater length in Self-study materials and for AI tutors in AI tutoring.

Learning from real use

Adults can keep learning from real attempts. Structured debriefs (after-action reviews) improved performance over controls with a medium-large effect, similar for teams and individuals and in real or simulated settings (Tannenbaum & Cerasoli, 2013). Their effectiveness is bolstered by alignment, meaning a debrief matched to the level and content of the task, and possibly by facilitation and structure (Tannenbaum & Cerasoli, 2013). Feedback on performance helps on average, but over a third of feedback interventions made performance worse, and effects shrink when feedback directs attention to the self rather than the task (Kluger & DeNisi, 1996).

Written references for use

A reference is consulted during work rather than studied beforehand. The best-known design idea is the minimal manual, built around real tasks and error recovery. In a controlled experiment, learners using an action-oriented minimal manual completed about 50 percent more tasks successfully and consulted the manual less than learners using a conventional one (van der Meij & Lazonder, 1993).

Usage examples help in technical documentation: in a controlled study of 26 engineers, API documentation without usage examples caused problems with data types, formats, headers and request bodies (Sohan et al., 2017). Checklist-style job aids raised accuracy and lowered strain in a randomised experiment, but slowed task completion (Fletcher et al., 2018).

Questions, controversies, and new developments

Do tasks from the learner’s own work help?

Weak evidence

The case is plausible but not shown for adults. The studies of personalised tasks use school pupils, and the 4C/ID comparison is against unspecified controls (Costa et al., 2022; Bernacki & Walkington, 2018). The wording-personalisation result is a null or negative one, and it tests pronouns rather than own-task content (Almeida et al., 2024).

Does comparing methods build understanding?

Contested

In one classroom study, comparing solution methods raised procedural flexibility and procedural knowledge but not conceptual knowledge (Rittle-Johnson et al., 2017). It helped conceptual and procedural knowledge mainly for pupils who already knew a method, and it helped flexibility across pupils (Rittle-Johnson et al., 2017). A randomised trial of a comparison-based supplementary curriculum did not raise algebra knowledge; teachers used it little, and greater use helped only procedural knowledge (Rittle-Johnson et al., 2017). The studies of conceptual and procedural knowledge cited here are of children learning mathematics, and the latent-variable link between the two is correlational (Schneider et al., 2011).

Is a prescribed sequence good for self-directed adults?

Weak evidence

Whole-task designs recommend simple-to-complex ordering (Daniel et al., 2018). But in self-paced distance courses (n = 540 college students), following the designed sequence correlated negatively with successful completion; the study is correlational and shows no causal effect (Lim, 2016). Its author argues that self-paced courses should give learners control of sequence, tied to adult learners’ prior knowledge (Lim, 2016).

Do the checks that are easy to collect predict later use?

Weak evidence

Reaction measures are the most used evaluation in organisations (78 percent against 9 percent for behaviour), probably because they are easy to collect (Arthur et al., 2003). Arthur and colleagues state that reactions show little or no relationship to the other three levels of evaluation, citing Alliger and colleagues for weak intercorrelations among the criteria (Arthur et al., 2003). An evidence summary reports that reactions are only modestly related to immediate knowledge and unrelated to delayed knowledge, and that utility reactions modestly relate to transfer while affective reactions do not (Quality Improvement Center for Workforce Development, 2022). When learning criteria were compared with behavioural ones, effect sizes fell substantially (Grossman & Salas, 2011).

Do tests other than closed-book recall help or predict use?

Contested

Practice tests often show no positive transfer after correcting for publication bias, unless response congruency, elaborated retrieval or good initial test performance is present (Pan & Rickard, 2018). The self-explanation meta-analysis hypothesised a larger effect on transfer than on recall, but found g = .534 for transfer and g = .498 for recall, a non-significant difference (Bisra et al., 2018). In a reanalysis, direct teaching gave g = 0.84 against 0.48 for indirect teaching (Kobayashi, 2019). A systematic review of teach-back in chronic-disease education found improved knowledge and broadly positive outcomes, but one study found no quality-of-life gain and larger trials are needed (Dinh et al., 2016). These are results about learning, not about later use.

Do people use references, and how well do they find what they need?

Weak evidence

The minimal manual advantage is disputed. The independent replication was small (13 against 12 subjects) and its training-time advantage was not significant, unlike the 40 percent claim of the original (van der Meij & Lazonder, 1993). Critics note that the original comparison may be confounded, because the minimal manual was based on the control manual and may simply be a better second effort (van der Meij & Lazonder, 1993). Reviewers found the empirical base thin: three of its four features (brevity, real tasks, error recovery) are uncontroversial, but the effect of guided exploration is confounded with other variables (Williams & Farkas, 1992). In interviews, users avoided both paper and online help, citing trouble finding search terms and shallow explanations, and often solved problems by asking a colleague or experimenting (Novick & Ward, 2006). The checklist experiment used students on a lab task, and its authors note that aids have shown mixed results in transfer of training (Fletcher et al., 2018).

Can AI-assembled references be trusted?

Weak evidence

Studies of AI-generated reference content find three failures. ChatGPT-generated literature reviews contained many fabricated citations, and many real citations carried substantive errors; GPT-4 did better than GPT-3.5 but problems remained (Walters & Wilder, 2023). A human audit of four generative search engines found that many cited sentences were not fully supported by their citations, so a real citation does not mean the claim is supported (Liu et al., 2023). And errors propagate from upstream research: ChatGPT did not flag retractions or errors in papers it evaluated (University of Sheffield, 2025), and ChatGPT-5.0 used most retracted stem cell articles in question answering, with fewer than half of those answers carrying a retraction notice (Zhang et al., 2025). One proposed practice is to validate references against scholarly databases such as Crossref, Semantic Scholar and OpenAlex, since existing reference managers do not check whether citations are real (Abbonato, 2026). The figures describe models of 2023 to 2025.

Broader connections

Transfer connects instruction to problem solving, since flexible use of a skill is largely the ability to solve a new problem with it, and to memory, since what cannot be retrieved cannot be used. The designs above are also the building blocks of the study materials in Self-study materials and the guidance an AI tutor gives. Understanding as an outcome of explanation is covered in Explanation and understanding.

Further reading

Known gaps

  • No study was found that compares an adult’s own-work example with a generic example on later use of the topic; the evidence on personalised tasks comes from school pupils.
  • The main meta-analyses of how reactions and learning measures relate to on-the-job behaviour (Alliger et al. 1997, Sitzmann et al. 2008, Blume et al. 2010) could not be read; the article relies on later descriptions of them.
  • No source tested whether explaining in one’s own words, teach-back or open-book checks predict later real use.
  • Evidence on headings, signalling and information mapping for looking things up in a reference was not found; the reference-use studies are mostly 1990s software manuals.
  • Personal knowledge records, and how a reference or returning tutor supports reflection on real use, were not researched beyond debriefs and feedback.
  • Prerequisite ordering of modules, the productive-failure meta-analysis, the Dochy problem-based-learning meta-analysis and the Lin interest-personalisation meta-analysis were not read.
  • The terminology of the field (transfer of training, far transfer, adaptive expertise, performance support) was not researched as a question of its own.

References