Skip to content

Problem-based Learning

Principle · All principles
Evidence · 51 claims (17 for, 31 mixed, 3 against) · 47 studies (9 causal, 9 quant-synthesis, 9 review, 8 qualitative, 6 theoretical, 4 associational, 2 design), q1–q4 · 8 of 47 report an effect size · 46 claims rest on one study

Conditional relationship

Problem-based learning (PBL) organizes learning around ill-structured or authentic problems that learners investigate, usually in small groups with a facilitator. Read the evidence in two layers.

As a curriculum format, its average effect is small and unreliable. The best-documented synthesis in the wiki, a re-analysis of a PBL meta-analysis of 353 outcomes, reports a modest overall effect (g = 0.27) with large heterogeneity, individual outcomes from g = −1.26 to g = 1.91, publication bias that would shrink the effect to about g = 0.10, and no tutor-background category that predicts learning. So "use PBL" is not, by itself, a decision the evidence supports or rules out: what varies between implementations matters more than the label.

Within a problem-centred design, guidance is the better-supported lever. For a learner who is a novice on the target relation, leaving essential content to be discovered without support learns less than explicit instruction, while discovery with feedback, worked examples or scaffolding does better than other instruction. Starting with a problem before instruction is a narrower configuration (productive failure): it has meta-analytic support for conceptual knowledge and transfer when instruction follows, no procedural advantage, a reversal for second- to fifth-grade child learners, and is reported most reliably in well-structured domains. It is not evidence that open problems help novices learn content on their own.

The relationship is therefore conditional on the learner's starting knowledge of the target content, on the guidance and consolidation provided, on the outcome chosen (conceptual, transfer, procedural, product) and on the horizon. The pattern specifies a reusable design policy; its local diagnostic branches remain proposals.

Fitting the design to a situation

Ask for four facts when the brief omits them: what do learners already know of the content the problem needs?, is the outcome basic knowledge, or understanding and transfer?, how old are the learners?, and will explicit instruction consolidate the problem work?

If the brief says… Then change… Basis
The plan is to convert a whole course or curriculum to PBL and expects a clear gain Promise no gain from the label; the average effect is small (g = 0.27), ranges from strongly negative to strongly positive, and may be overstated by publication bias. Specify the guidance and consolidation and evaluate that configuration, recording how faithfully each group's sessions follow it The overall PBL effect is modest with large heterogeneity [~M]; Publication bias in the same data [~M]; An older review's equivocal results on knowledge and clinical reasoning, attributed to how faithfully programmes implement PBL [~W] (second-hand, in an essay)
The outcome that matters is basic factual knowledge on an exam soon after the unit (for example, a board or licensing exam) Do not rely on problem work alone for it; add explicit teaching and practice of the knowledge base, timed for the exam. If retention months later matters too, measure it separately: the two horizons have pointed different ways PBL: null or negative on basic knowledge, positive on understanding and transfer [~M]; Traditional instruction did better for short-term retention on standardized board exams [-W]; PBL did better for long-term retention [+W] (both from one meta-synthesis of meta-analyses, no effect sizes in the text available)
The outcome is conceptual understanding and transfer Use a problem-centred design with consolidating instruction afterwards PBL: positive on understanding and transfer [+M]; Problem-first improves conceptual learning when instruction follows [+M]
Learners are novices on the content the problem needs Make that content explicit (explanation, worked examples) before or alongside the problem; leaving it to be found unsupported learns less Minimal guidance is less effective for novices [-M]; Worked examples reduce novice search [+M]
Learners already have strong knowledge of the domain Lower the guidance and give more open problems, checking on an unaided task first The advantage of guidance recedes with high prior knowledge [~W] (a review's argument)
Learners are in the early primary years and the plan is problem-first Teach first and use the problem as application PS-I effects are negative for 2nd–5th graders and rise with age [-M]
Facilitators are not content experts (or the brief insists they must be) Do not choose tutors on content background alone; prepare every facilitator in the guidance and consolidation moves Tutor background did not predict learning [~M]
Small groups, each with a facilitator Have facilitators raise support after a failed step and lower it after success, rather than give fixed hints Contingent scaffolding improves learning [+M] (one-to-one tutoring, not PBL groups)
Tutorial groups are large Split them into small or medium groups for the problem work; do not run the self-directed phase as one large group Small and medium tutorial groups gained self-directed-learning readiness; large groups declined [~W] (one pre–post study reported in a review; a self-report scale, not achievement)
To save time, learners will be handed the learning issues for each problem Let each group generate its own learning issues from the problem, with the facilitator checking that they reach the target content Groups that generated their own learning issues studied more materials and spent more time on self-study [+W] (one experiment reported in a review; study behaviour, no achievement outcome)
Content is taught in lectures running alongside the problem sessions Write problems the lectured content informs but does not settle, and ask for questions and challenges rather than a summary of the lecture More lecture coverage went with less critical discussion in PBL sessions, but more willingness to speak [~W] (one qualitative cross-curriculum study)
Groups will work together over several weeks (online or face to face) Set up how the group plans, monitors and shares work in the first sessions, since groups' regulation settled early and rarely changed; online, re-launch the group's script part-way through Most groups settled into stable regulatory patterns early, with little change [~W] (one sequence-analysis study of collaborative groups); Online PBL groups drifted from the scripted sequence to social bonding [~W]; the early set-up and re-launch are untested proposals
AI tools (a generative chatbot, an adaptive tutor) will be added to the problem work Embed the tool in facilitator-guided problem work rather than leave it as an open answer source; check that learners can already use the tool fluently; evaluate the AI-supported configuration as a whole, on individual outcomes Benefits of generative AI in medical case- and problem-based learning appeared when it was embedded in faculty-guided work; in one pilot, open access to a language model did not improve accuracy and disrupted teamwork [~W] (a scoping review); AI-supported PBL gave high normalised gains against a control group's low gains [~W] (one quasi-experiment, which cannot separate the AI from the PBL); Years of coding experience went with higher scores in AI-supported PBL [~W] (exploratory, one course); a review of 50 studies lists as themes that AI supports educator–student collaboration through feedback and scaffolding [~W] and adaptive, personalised learning linked to engagement and problem solving [~W], with no comparisons on the claim pages; Adult learners expect AI to activate background knowledge and sharpen their questioning [+W] (20 interviews; expectations, not outcomes)
The goal is application of a skill, and a participatory simulation is an option Expect the simulation to give larger application gains and PBL larger knowledge gains, in one trial with 54 seventh and eighth graders; choose by the outcome that matters Simulation gave larger application gains [~W]; PBL gave larger knowledge gains [~W]
The learners are adult professionals in a preparation programme (for example school leaders) Tie the problems to the learners' own practice through fieldwork, not only to classroom cases Problem-based learning connecting coursework and practice [+W]
Collaboration skills are a goal Ask learners to reflect in writing on how the group worked, and prompt for thinking processes, which learners rarely mention unprompted PBL students named valuing others' perspectives and interdependence [+W]; Students rarely referred to higher-level processing benefits [~W]
Every problem product will be graded Keep the problem phase low-stakes and grade the individual consolidation and transfer tasks untested proposal

Observation, state and explanation

Record what the learner does with a stated problem under stated conditions: what they identify as the question, what they say they need to know, which resources they consult, which solutions they generate, what help they receive and from whom, and what they produce. "Ready for open problems" is an inferred state, not an observation. Activity, talk volume and enthusiasm are observations of participation; none directly measures whether the target concept was learned. Engagement is reported higher during PBL than during traditional instruction [+W] (second-hand, in a scaffolding framework), so expect it, and do not read it as learning. In a group, a shared product does not show what each member can do.

Activation of prior knowledge, noticing gaps and comparing one's own solutions with the canonical one are explanatory hypotheses offered by the claims below; they have not been observed directly in the studies the wiki records.

Response Plausible explanations Observation that could change the interpretation
Novice on the target content, given an open problem to learn it from Missing the content the problem needs; problem too hard to generate partial solutions; access or language demand Ask for the key relation directly in a short probe; if absent, teach it (explanation, worked example) before or alongside the problem and observe whether the problem work then uses it.
Cannot start: no question framed, no learning issues listed Missing task-specific knowledge to enter the problem; problem framed in unfamiliar language or representation; unclear what a product should look like Ask for a first decision with a short prompt in a familiar representation; show the form of a finished product without its content; record whether either releases a start.
Busy and engaged, but the product shows little use of the target concept Productive generation that awaits consolidation; floundering search without schema; target concept not actually needed by the problem as posed Ask the learner to name the idea their solution depends on; check whether a correct solution to the problem as written requires the target concept at all.
Group discussion is fluent explanation-building, with little checking against the problem's data The group is theorising ahead of its evidence; members are reasoning privately and the talk shows only part of it; the problem gives too few data to check against Ask which facts in the problem the explanation does not yet account for, and ask each member privately what they are unsure of. In one tutorial group, discussion during problem analysis was mainly theory building, with less data exploration and meta-reasoning [~W] (one qualitative study).
Group product strong, individual explanation weak Work carried by one or two members; learning distributed but not yet individually consolidated; explanation probe harder than the task Ask each learner for an individual explanation or a changed-condition problem under matched conditions; record who did what during the problem phase.
Generates several flawed solutions, then learns quickly from instruction Problem-first preparation, as productive-failure accounts propose; instruction alone would have sufficed; extra time on task Only a comparison with an instruction-first condition of matched time can attribute the gain; a single learner's trajectory cannot.
Succeeds on a conceptual probe, fails a routine procedure (or the reverse) Different outcomes from the same sequence; procedure not practised; probe misaligned with what was taught Score procedural and conceptual items separately, at the same horizon, and preserve both.

These rows are proposals, untested with learners. A probe can shift confidence among the explanations; it does not identify a cause on its own. The probe itself is a learning exposure and should be recorded as one.

Evidence and qualifications

  • The overall PBL effect in a tutor-background meta-analysis is modest, with large heterogeneity [~M]: a re-analysis (Walker & Leary, 2023) of the Leary et al. (2013) data, 353 outcomes, reports g = 0.27 with I² = 80% of variability attributable to heterogeneity; individual outcomes range from g = −1.26 to g = 1.91 [~M]. Not yet checked against its source. It does not predict the effect of any one PBL implementation, and the claim pages do not report learners, comparators or horizons by outcome.
  • Publication bias in the same data [~M]: a funnel plot and Egger's test indicate publication bias, and trim-and-fill drops the overall effect to g = 0.103. Treat the average as possibly overstated.
  • Tutor background does not predict learning [~M]: in a meta-regression with content novices as the reference, no tutor-expertise category predicted student learning (R² = .01, p = .14), and the largest subgroup difference was not significant [~M]. A nonsignificant moderator is not evidence that facilitation does not matter; it says tutors' credentials alone did not explain the variation.
  • A small meta-analysis reports a large pooled effect on achievement [+W]: five studies, random-effects ES = 1.560 (95% CI 0.768–2.353), heterogeneous; the claim page records that the article describes its z-test as not significant (z = 3.859; p > .05), which is internally inconsistent. Do not weigh it against the 353-outcome synthesis.
  • Minimal guidance is less effective for novices than explicit instruction [-S]: a meta-analysis of 580 comparisons found explicit instruction outperformed unassisted discovery (d = 0.38 favouring explicit instruction), while 360 comparisons found enhanced or assisted discovery (feedback, worked examples, scaffolding, elicited explanations) outperformed other instruction. An experiment with 112 third- and fourth-grade novice children (the claim page does not say how they were assigned) found many more mastered the control-of-variables procedure under direct instruction, and those taught directly did as well on a later transfer task as the few who discovered it. The Klahr & Nigam entry passes the judge; the registry attaches the wrong abstract to the meta-analysis (recorded as a registry problem), and the narrative review is unsettled. This bounds the principle: it predicts poor learning when PBL means leaving novices to discover essential content unsupported, not that guided problem work fails. The claim page does not break the meta-analytic result down by prior knowledge.
  • More teacher guidance within inquiry goes with larger advantages over traditional instruction [+M]: a review of inquiry-learning syntheses reports one science meta-analysis with d = 0.25 for minimal-guidance inquiry over traditional instruction against d = 0.65 for teacher-directed inquiry, and a second meta-analysis finding only g = 0.15 for more over less guided inquiry. Both second-hand, about inquiry rather than PBL specifically; they point the same way as the default (guide the problem work) but disagree on how much it buys.
  • Practising whole complex tasks is argued to favour transfer [+W]: a training-design argument, the rationale for posing a whole problem rather than its parts; no study comparing whole-task with part-task practice is recorded on that claim, and the pattern records one experiment in which absolute novices did better with isolated parts first.
  • Productive Failure Improves Conceptual Learning [~S]: a three-level meta-analysis (synthesis-mixed: experimental and quasi-experimental comparisons, 53 studies, 166 comparisons) set problem solving followed by instruction against the same instruction taught first. It favoured problem-first on conceptual knowledge and transfer (g = 0.36) and found no difference on procedural knowledge (g = −0.03); the advantage was larger for implementations following productive-failure design criteria (generating multiple solutions, group work, instruction building on student solutions) and reversed for second to fifth graders and for domain-general skills. One entry passes the judge on its abstract; the synthesis entry is unsettled on its abstract. It does not predict an advantage for problem-first work with no subsequent instruction, for young children, or on procedural fluency, and the claim page reports no delayed-horizon breakdown. Marked ~ because its support for problem-first work is conditional on instruction following and on learner age.
  • Contingent scaffolding improves learning more than fixed or absent support. [+M]: one-to-one tutoring studies in which support was adjusted to each learner response outperformed fixed, moderate or no support on immediate and one-month tests of long division (N = 8 per condition), and interactive tutoring gave similar immediate learning but better transfer; a dynamic-assessment synthesis ranked scaffolding above coaching but below explicit strategy training. None of its three judged entries passes; they could not be confirmed on their abstracts. Its evidence comes mostly from one-to-one tutoring, not from facilitated PBL groups, so it supplies a candidate facilitation policy rather than a tested PBL result.
  • Case-based learning improves exam performance [~M]: in one community-college biology cohort (n = 56; a non-randomized within-cohort comparison of matched topics), case-taught topics scored higher on course exams; a systematic review of 104 papers in health professional education (it passes the judge) concluded that students enjoy case-based learning but the evidence on learning compared with other activities is inconclusive, and that any benefit of small-group case work may come from the group work rather than the cases. It does not establish an advantage for authentic-problem formats in general and does not report delayed outcomes.

Retain learner age and starting knowledge, the exact comparator (instruction-first, lecture, unassisted discovery), the amount and timing of guidance, time on task, outcome type and horizon when transporting these results. A pooled g over unlike implementations is not a forecast for one class or one learner, and a nonsignificant procedural difference does not establish equivalence.

The evidence by question

Each question below links a general claim that gathers the studies behind it into one table, one row per distinct study, with each study's design and effect as recorded and no averaged verdict (eval/synthesis/problem-based-learning.json, scripts/synthesize_claim.py). The count is of distinct studies across the claim and the narrower claims that bear on it.

Further evidence, not yet read against this model

Claims this page cited before it was rewritten as a conditional model. They are evidence about the relationship, but none has yet been re-read against the model above; each says how far its own sources have been checked.

On 2026-10-19 (wave 2) the other claims linked here were read and moved into the model: worked examples (already in the novice row), whole-task practice and teacher-guided inquiry (evidence and qualifications), the short- and long-term retention pair (the basic-knowledge row), the equivocal-evidence essay (the curriculum row), engagement and theory-building discussion (observation), the evidence-gaps review (by question), and the rows on group size, learning issues, lectures alongside PBL, groups working over several weeks, and AI tools. - Spontaneous responsiveness to real-world events in PBL can deepen student-directed inquiry beyond what designed curriculum achieves [+W] - In one extreme case, two sixth-graders with a choice-rich, ungraded STEAM infrastructure extended a game-design challenge into a year-long self-directed project well beyond its designed scope [+W] - Students rated real-world projects and discussions with structured or guiding questions as the most beneficial learner-to-content strategies [+W] - Talladega County Schools' graduation rate rose from just above 70 percent in 2008 to 93 percent after its transition to project-based learning supported by digital technologies [+W]

Objective and learner-valued goal

Ask what the learner wants from the problem, and why, separately from the designer's objective. A problem chosen for its authenticity to the designer may not be one the learner cares about, and a learner's interest in solving the problem may diverge from the designer's interest in a concept the problem is meant to carry. Agreement, divergence and uncertainty are observations; engagement during the problem phase does not establish that the learner values the target capability. Choose a representative instrument, criterion, permitted resources and horizon for each outcome that matters: conceptual understanding, transfer, procedural fluency and the quality of the product are different claims.

For example, a nursing student may value making a safe decision for the patient in a case, while the designer's objective is that the student can explain the physiological principle the case was built around. A correct group care plan cannot establish either the individual decision or the explanation. The pattern shows how to record both aims and test the one each party cares about.

What would revise this model?

The problem-first expectation should weaken if comparable learners, with comparable consolidation instruction and time, fail to show a conceptual or transfer advantage over instruction-first sequences on aligned measures, or if the advantage disappears at a delayed horizon. If younger learners, domain-general skills or other populations repeatedly show the reversal recorded in the synthesis, narrow the principle to the populations where it holds. If matched-time comparisons attribute PBL gains to group work, feedback or extra time rather than to the problem-first order, revise the explanation. Do not protect the model by relabelling every failure as poor facilitation or insufficient authenticity after the fact.

An engaging problem phase, an immediate conceptual gain, delayed retention, transfer to practice and the learner's valued use are separate claims. The evidence recorded here does not establish an optimal amount of guidance, a readiness threshold for open problems, or the effect of whole PBL curricula against conventional instruction.

At the curriculum level, the expectation of a small and variable average should change if bias-corrected syntheses with stated comparators and horizons show a consistent effect, or if moderators (guidance, consolidation, prior knowledge, outcome type) are found that explain the variation.

Claims

Examples

  • Community issue investigation: Learners analyze a local problem, gather evidence, and propose responses.
  • Case-based team problem solving: Small groups work through a realistic professional dilemma with incomplete information.
  • Design challenge: Learners create and defend a solution to a practical constraint-based problem.
  • Cross-disciplinary PBL module: Learners integrate literacy, numeracy, and research skills to address a shared problem.

Key Sources

  • Marra, R. M., Jonassen, D. H., Palmer, B., & Luft, S. (2014). Why problem-based learning works: Theoretical foundations. Journal on Excellence in College Teaching, 25(3-4), 221-238.
  • Thorndahl, K., & Stentoft, D. (2020). Thinking critically about critical thinking and problem-based learning in higher education: A scoping review. The Interdisciplinary Journal of Problem-Based Learning, 14(1). https://doi.org/10.14434/ijpbl.v14i1.28773
  • Lin, L. F. (2017). Impacts of the problem-based learning pedagogy on English learners' reading comprehension, strategy use, and active learning attitudes. Journal of Education and Training Studies, 5(6), 109-125. https://doi.org/10.11114/jets.v5i6.2320
  • Savery, J. R. (2019). Overview of problem-based learning: Definitions and distinctions. In R. West (Ed.), Foundations of Learning and Instructional Design Technology. EdTech Books. https://edtechbooks.org/lidtfoundations/overview_of_problem_based_learning
  • Molenda, Michael. (2002). A New Framework for Teaching in the Cognitive Domain. ERIC Digest. https://eric.ed.gov/?id=ED470983
  • Lundberg, M., & Rasmussen, J. (2018). Foundational Principles and Practices to Consider in Assessing Maker Education. i-manager's Journal of Educational Technology, Vol. 14 No. 4, January - March 2018. https://eric.ed.gov/?id=EJ1179517
Page metadata
Field Value
type principle
id problem-based-learning
aliases real-world-problem-solving-maker-context
title Problem-based Learning
description For learners who must apply knowledge to ill-structured problems, organizing work around problems has a small, highly variable and possibly inflated average effect; what is better supported is guidance for novices on the target content, with a problem-first phase as a narrower option followed by instruction.
status review
moved_from pattern/problem-centered-task-level-progression-pattern
canonical True
generated by claude/unspecified on 2026-10-19
trust unverified — no one has confirmed this page's content
sources 4 — thorndahl-2020, lin-2017, molenda-2002, lundberg-2018

Frontmatter is the page's machine-readable record — see the Schema & Guide.