On this page
Most textbooks and coaching modules follow one order: here is the formula, here is a solved example, now practise. A different idea says you should first wrestle with a problem you have not been taught to solve, fail in useful ways, and only then get the formal method. This is called productive failure. A 2021 meta-analysis by Tanmay Sinha and Manu Kapur pulled together the evidence. Here is what it found, and where it does not apply.
What question did the study ask?
The paper, When Problem Solving Followed by Instruction Works: Evidence for Productive Failure in the Review of Educational Research, compared two orders of learning a new concept:
PS-I: problem solving first, followed by instruction.
I-PS: instruction first, followed by problem solving.
Productive failure is a stricter version of PS-I, developed in Kapur's earlier work such as his 2016 paper in Educational Psychologist. The problem-solving phase is deliberately designed so that students generate several partial solutions but usually do not reach the correct method. A teacher then builds on what students tried to teach the target concept.
What was studied?
According to the open-access full text, the review covered 53 studies with 166 comparisons: 81 experimental and 85 quasi-experimental. Most studies were in Europe, North America and Asia, and most taught maths concepts to students in grades 6 to 10 or to undergraduates, with some work in physics, chemistry, psychology and general reasoning skills.
The authors also coded each PS-I study for how closely it followed the design rules of productive failure, such as whether students generated multiple solution attempts, worked in groups, and whether the teacher's later explanation built on the students' own attempts.
What did it find?
The paper reports effect sizes as Hedges' g, a standardised difference between groups:
Conceptual knowledge and transfer: g = 0.36 in favour of problem solving first (95% confidence interval 0.20 to 0.51). That is a moderate effect.
Procedural knowledge (carrying out the steps): g = -0.03 across 51 comparisons, which is effectively no difference. Trying first did not hurt skill at doing the procedure.
Fidelity mattered: studies that followed productive failure principles closely had larger effects, between 0.37 and 0.58. Four design features were linked to bigger effects: evidence that students generated multiple solution attempts, group work, instruction that built on student solutions, and discussion-based teaching afterwards.
Age mattered: for students in grades 2 to 5 the effect was negative (g = -0.09), favouring instruction first. For grades 6 to 10 it was 0.50.
Domain-general skills, such as controlling variables in an experiment, also favoured instruction first.
After a statistical correction for publication bias, the estimated effect rose to 0.87. Treat that figure with caution: corrections like this rely on statistical assumptions, so the 0.36 measured across the studies is the safer number to remember.
What does it not show?
It is not "discovery learning". Every study included instruction afterwards. Struggling alone with no teaching to follow is a different thing.
The problem design matters. Low-fidelity versions gave smaller effects. Any random hard question is not the same as a designed productive failure task.
Effects varied a lot across studies, and few were in India or in exam-coaching settings.
It conflicts with some worked-example research, which finds that novices learn more from studying solutions first. Our post on worked examples vs solving from scratch covers that side.
A sample productive failure task
Here is an illustration using sample numbers, before you have been taught standard deviation. Two batters score these runs in three innings:
Batter A: 40, 50, 60
Batter B: 10, 50, 90
Question: who is more consistent? Invent a way to measure it. Students usually try the range (A: 20, B: 80), or the average distance from the mean. Both averages are 50, so the distances are 10, 0, 10 for A (average 20 divided by 3, about 6.7) and 40, 0, 40 for B (about 26.7).
Then the teacher shows the formal method: square the deviations and average them. A gets (100 + 0 + 100) divided by 3, about 66.7, and B gets (1600 + 0 + 1600) divided by 3, about 1066.7. Taking square roots gives standard deviations of about 8.2 and 32.7. Because you have already invented the range and average distance, you can see why the formal method is built the way it is.
What does it mean for students in India?
Before a new chapter, spend 15 minutes on one problem from it before the lecture. Write down every approach you try.
Compare attempts with a friend. Group work was one of the features linked to larger effects.
Then learn the method, and look for where your attempts broke. That comparison is the point. Our post on why mistakes help you learn explains the error-correction side.
Parents of younger children: for primary-school maths, clear teaching first is what this review supports.
After your attempt, you need a clear explanation to compare it with. The Fermi Concept Library covers every concept in physics, chemistry, biology, mathematics and computing, so you can read the formal idea right after your own attempt.
How to read a study like this
In a meta-analysis, check the subgroups as well as the headline. Here the average effect was positive, but one subgroup (young children) went the other way. A single average can hide groups for whom the advice is wrong.
Next step: pick tomorrow's new topic, find one problem from it, and give yourself 15 minutes to attempt it before the class.
