In 1975 a researcher named Joan Rubin published a paper in TESOL Quarterly with a question in the title that sounds almost naive: what can the good language learner teach us? The logic behind it was simple and slightly ruthless. Instead of starting from theory and asking what should work, Rubin started from results and asked what the successful students were already doing. If you could describe the behavior of people who got fluent, you could hand the description to everyone else.

That single question launched a research program that is now fifty years old, has generated thousands of papers, several competing taxonomies, at least three meta-analyses, and one long argument about whether "learning strategies" is even a coherent thing to study. It also produced one of the most useful and least glamorous findings in the field: the difference between learners who succeed and learners who stall is mostly not talent, and mostly not materials. It is a small set of decisions about what to do when you do not understand something.

This is the state of the evidence, including the parts that do not flatter the idea.

The List From 1975

Rubin built her paper from observation of good learners in real classrooms rather than from laboratory tasks, and the traits she described have held up well enough that later researchers kept rediscovering them. The good learner, in her account, is a willing and accurate guesser who treats ambiguity as information rather than as failure. They have a strong drive to communicate, which means they will accept sounding worse in order to say something now. They are often uninhibited and willing to appear foolish. They attend to form as well as meaning, noticing patterns in what they hear. They practice and monitor their own speech and other people's, and they treat every conversation as potential input rather than as a test.

Read that list again and notice what is missing. There is no mention of memorizing word lists, no mention of apps, no mention of studying grammar tables in order. The behaviors Rubin isolated are almost all about how a learner handles not knowing something. That is the through-line of the entire literature, and it is why strategy research is more useful to intermediate learners than to beginners: beginners spend most of their time simply receiving, while everyone past the first few hundred words is constantly making decisions about what to do with a gap.

What the Ontario Team Added

Three years later, a team at the Ontario Institute for Studies in Education — Naiman, Fröhlich, Stern and Todesco — published The Good Language Learner, which remains the reference point for this kind of work. They interviewed successful learners and their teachers and condensed what they found into five broad strategies that read like a job description:

  • An active task approach — the learner treats the language as a project with tasks, not as a subject to be absorbed.
  • Realization of language as a system — treating grammar as a structure to be inferred rather than a set of rules to be obeyed.
  • Realization of language as communication — using it with actual people for actual purposes, early and badly.
  • Management of affective demands — handling the embarrassment, the boredom and the fear of sounding stupid.
  • Monitoring of performance — listening to your own output critically enough to notice what went wrong.

Two of those five are emotional, not cognitive. That was radical in 1978 and it is still the part most study advice ignores. Rubin's "uninhibited" and Naiman's "affective demands" are describing the same mechanism from different angles: language learning produces a continuous stream of small public failures, and the learners who progress are the ones who have found a way to not treat each failure as evidence about themselves.

Two Taxonomies, One Problem

In 1990 the field tried to organize itself, twice, in the same year. O'Malley and Chamot proposed three categories: metacognitive strategies (planning, monitoring, evaluating your own learning), cognitive strategies (manipulating the material — summarizing, inferring, elaborating), and social/affective strategies (using other people, and managing your own feelings). Rebecca Oxford, in Language Learning Strategies: What Every Teacher Should Know, proposed six: memory, cognitive, compensation, metacognitive, affective and social. Her book also shipped the instrument that dominated the next fifteen years of research — the Strategy Inventory for Language Learning, or SILL, a questionnaire that asked learners to rate how often they did things like "I try to talk like native speakers."

The SILL made the field measurable, and measurement produced a mountain of correlational studies showing that better learners report using more strategies. That result is real and it is also nearly uninterpretable, because it has the same shape as the finding that good readers report reading more. The question strategy research had to answer was not whether successful learners use strategies. It was whether teaching strategies to unsuccessful learners makes them better.

Oxford's own taxonomy later drew criticism for internal inconsistency — memory strategies are arguably a subtype of cognitive strategies, and compensation strategies are about using language rather than learning it. O'Malley and Chamot's third category was criticized as a catch-all. Ernesto Macaro spent years cataloguing the deeper problem: researchers could not agree on whether strategies are mental operations or observable behaviors, whether they are knowledge, intention or action, or whether they transfer across tasks at all. By the late 1990s some researchers had given up on the term entirely and replaced it with self-regulation, the psychological construct for how learners set goals, monitor, and adjust.

The Meta-Analytic Ledger

This is where the argument gets settled, or at least bounded. In 2011 Luke Plonsky published a meta-analysis of second language strategy instruction in Language Learning that collected the whole population of studies and coded them for moderators.

"Effect sizes were calculated for 61 primary studies, contributing a total of 95 unique samples... The findings indicate a small to medium overall effect of SI (d = 0.49). Variables found to moderate its effectiveness include type and number of strategies, learning context (second vs. foreign language), and length of intervention."

A d of 0.49 is not a miracle and it is not nothing. Converted to a percentile, a learner at the median of the control group would sit at roughly the 69th percentile of that group after receiving strategy instruction. For comparison, that is in the same range as many well-regarded classroom interventions, and it comes from teaching students how to approach a task rather than giving them more vocabulary.

Plonsky's moderator findings are the practically important part. The effect depended on the type and number of strategies taught, on whether learners were studying the language in a country where it is spoken (second language) or in a classroom far from it (foreign language), and on the length of the intervention. Short workshops that name a strategy and move on produce small effects. Longer instruction that lets learners apply, fail, and reapply produces larger ones. A strategy you have heard of is not a strategy you own.

Three further data points sharpen the picture. Dalman and Plonsky (2022) ran the same analysis restricted to listening strategy instruction and found a medium effect of d = 0.69 across 45 primary studies and 51 independent samples — larger than the general figure, which makes sense given that listening is the skill learners most often describe as an undifferentiated wall of sound. A 2020 meta-analysis of 90 meta-analyses, covering 3,496 studies and roughly 700,000 students, reported that language learning strategies had a medium overall impact and the largest impact on speaking of any teaching practice examined, at d = 0.90. And Plonsky's own replication work with Oswald on study quality found that the strategy literature is unusually prone to small samples and under-reported statistics, which means the true effect is probably somewhat smaller than the published number.

The Two Cleanest Experiments

Meta-analyses average over messy studies. Two individual experiments are worth knowing because they show the mechanism in a form you can copy.

Vandergrift and Tafaghodtari, 2010

This study took 106 students in six intact sections of French as a second language and split them into an experimental group of 59 and a control group of 47, taught by the same teacher, listening to the same texts the same number of times. The only difference was process. The experimental group was walked through the metacognitive cycle before, during and after listening: prediction and planning before the text, monitoring during it, problem solving when comprehension broke down, and evaluation after. The control group simply listened.

The experimental group significantly outperformed the control group on the final comprehension measure after controlling for initial differences. Two details matter more than the headline. First, the gains were largest for the less skilled listeners — the students who normally get nothing from listening practice were the ones who benefited most. Second, the researchers tracked metacognitive awareness with a questionnaire across the semester and found it changing, which is evidence that something was actually being learned about learning.

Graham and Macaro, 2008

Sixty-eight lower-intermediate learners of French in England received strategy instruction in listening, compared against a comparison group, with the intervention delivered at two levels of scaffolding. Both listening proficiency and self-efficacy improved. The self-efficacy result is the one to sit with. Learners who believe they are bad at listening tend to stop attending to the parts they miss, which produces worse comprehension, which confirms the belief. A strategy gives them something to do in the moment of failure, and doing something interrupts the loop.

The Case Against, Which Is Stronger Than Most Articles Admit

In 1993, Janie Rees-Miller published "A Critical Appraisal of Learner Training" in the same journal that had run Rubin's paper eighteen years earlier. Her objections were not that strategies do not exist but that the causal story was unproven. If good learners use more strategies because they are more proficient — because fluency gives them the spare attention to plan and monitor — then training strategies in isolation is teaching the symptoms. She also questioned whether strategies are separable from language knowledge at all, and whether "learner training" transfers beyond the specific task it was taught on.

Vann and Abraham (1990) supplied the empirical sting. They compared successful and unsuccessful learners and found that the unsuccessful learners also reported using strategies — in some cases more of them. What separated the groups was not quantity but fit: whether the strategy matched the task, and whether the learner could tell when it was not working and switch. Reporting that you "guess from context" is worthless if you guess from the wrong cue and never check.

So the honest summary is this. Strategies are real, teachable, and worth about half a standard deviation when taught properly over time; the effect is largest for listening and speaking; and strategy counts are close to meaningless. What predicts success is a small number of strategies used at the right moment, with monitoring that catches the failures.

What Actually Survives

Strip out everything that failed to replicate or could not be operationalized, and five things remain. Each one is a decision you make in the moment of not understanding.

1. Predict before you check

The metacognitive cycle that produced the 2010 result is not complicated. Before you reveal an answer — on a flashcard, in a listening exercise, in a sentence you are reading — commit to a guess and say why. The guess is the point, not the answer. This is the single highest-yield change you can make to a review session, because it converts passive recognition into a prediction with a stake in it.

2. Guess from the sentence, not the word

Rubin's "willing and accurate guesser" is really two skills: willingness, which is emotional, and accuracy, which is grammatical. Accurate guessing uses the words around the gap, the part of speech the gap must be, and the register of the text. Train it deliberately: cover a word in a sentence you already understand and reconstruct it before looking.

3. Monitor for the break, not for the error

Naiman's monitoring strategy is often taught as self-correction, which turns review into an anxiety exercise. The useful version is narrower: notice the exact moment comprehension stops. Was it a word, a contraction, a referent you lost two sentences ago, or a speed problem? The answer determines what to practice, and most learners never ask because they experience the failure as a fog rather than as a location.

4. Use compensation out loud

Compensation strategies — circumlocution, approximation, asking for a repair — are the ones that keep a conversation alive when your vocabulary runs out. They are also the ones learners skip, because using them requires admitting the gap in public. A useful drill: describe an object you do not know the word for, for thirty seconds, without stopping.

5. Regulate the feeling, not just the schedule

Both Rubin and Naiman put affect in the list and both modern taxonomies kept it. The evidence-based version is not "stay motivated"; it is reappraisal. Anxiety before speaking is arousal, and arousal that gets labeled as excitement tends to improve performance relative to arousal labeled as dread. The practical move is to name the physical state and re-label it before you open your mouth.

A Weekly Strategy Audit

Because the moderator that mattered most in the meta-analysis was length of intervention, the protocol below is built as a repeating cycle rather than a one-off lesson. Twenty minutes, once a week, alongside your normal reviews.

  1. Pick one task that keeps failing. Not "listening" — a specific task, like understanding fast dialogue in a show without subtitles.
  2. Name the strategy you will apply. One, from the five above. Write it as a sentence you could act on.
  3. Do the task cold and record what happened. Where exactly did it break?
  4. Do it again with the strategy. Predict first, guess from the sentence, monitor for the break.
  5. Compare. Did the strategy change the outcome, or did you just feel busier?
  6. Write one card. The card is not the vocabulary word; it is the cue. "Before replaying, predict the topic from the first three words."
  7. Schedule the same task next week with the same strategy, on new material. Transfer is the whole test.

The transfer step is where most strategy training dies. A strategy practiced once on one text is a party trick. The question that matters is whether you reach for it automatically six weeks later on material you have never seen.

Three Failure Modes

  • Strategy collecting. Reading about forty strategies and using none, because each one felt like progress. Plonsky's moderator on number of strategies is a warning here: piling on more is not the mechanism.
  • Strategy as avoidance. "I use metacognitive strategies" can mean planning elaborate study schedules instead of doing the listening. Planning feels productive and costs nothing.
  • Strategy without feedback. A guessing strategy with no check is just confident error. Every inference needs a resolution — look up the word after you guess, replay the segment after you predict.

What This Means for a Deck

Flashcards sit at an awkward angle to this literature. Spaced repetition is a scheduling technology, and strategy research is about moment-to-moment decisions, so a deck can either support strategies or quietly replace them. The deck that supports them looks different in three ways.

First, it makes you produce before it shows you. A card you recognize is a card that has not asked you to decide anything. Second, it keeps the sentence, not just the word, because the guessing strategy operates on context and a bare word gives context nowhere to live. Third, it surfaces the failure honestly — you mark the card wrong because you could not retrieve it, not because you feel shaky, since the feeling is exactly the signal that misleads people in both directions.

None of that is a strategy in itself. It is an environment where the strategies that survived the meta-analyses have somewhere to happen.

The Bottom Line

Rubin's question has an answer, and the answer is smaller and more specific than the self-help industry suggests. Successful learners are not doing forty things you are not doing. They are guessing more often, checking their guesses, monitoring where comprehension breaks, and managing the embarrassment of being wrong in public. Taught over months rather than minutes, those behaviors are worth roughly half a standard deviation — enough to move a stalled learner, not enough to replace the years of contact that fluency actually costs.

The uncomfortable corollary is that knowing the list changes nothing. Rees-Miller's critique applies to every reader of this article: the gap between describing a strategy and using it under pressure is the entire problem. Pick one. Apply it this week on material that is slightly too hard. Then check whether it actually helped, because the learners who improve are the ones who check.