EğitimGelişim

The Intermediate Plateau: Why Progress Stalls After the Beginner Phase and How to Break Through

A learner pauses during guitar practice by a bright window with a staircase behind, illustrating the intermediate plateau in skill learning

TL;DR. A plateau is not one thing. The research record describes at least six causes, and they call for opposite responses. Part is arithmetic: in Cambridge’s guided-learning-hour estimates, the step to A2 English costs about 95 hours at the midpoint and the step to C2 about 350, or 3.68 times as much for one more label. Part is automaticity: Ericsson puts the point where everyday skills become automatic at frequently less than 50 hours, after which experience alone stops helping. In a systematic review of 62 evaluations of physicians, 32 (52%) reported decreasing performance with more years in practice for all outcomes assessed, and only 2 reported performance increasing. Part is method: a study of 25,280 individual learning curves found a brief drop in performance at each switch to a better strategy. The CEO move is to decide whether this plateau deserves to be broken at all: stop, maintain or push. The student move is to change the design of practice, not the number of hours.

This article belongs to the CEOtudent learning-science cluster. The hub is the meta-skills that make every other skill easier to learn. Two companion pieces cover the ground this one deliberately does not repeat: what expertise research actually found about the 10,000-hour rule, and deliberate practice in the age of AI.

Where does the idea of a learning plateau come from?

One of the earliest formal studies of a plateau came from telegraphy. In 1899 William Bryan and Noble Harter published a study in Psychological Review of how operators learned to receive Morse code. Their curves came from individuals tested each week, confirmed by a consensus of opinion from about two hundred operators. The sending curve rose smoothly. The receiving curve did not: it climbed, flattened for months, then rose again.

Their explanation is still the most useful sentence written on the subject: “A plateau in the curve means that the lower-order habits are approaching their maximum development, but are not yet sufficiently automatic to leave the attention free to attack the higher-order habits.” A learner who still has to think about letters has no attention left for words.

They also noticed that the pattern was not special to telegraphy. On learning a foreign language they wrote of “the same rapid improvement at first, the same dispiriting level just below the ability to understand ordinary conversation.” On English composition, instructors at Indiana University reported rapid progress nearly up to the passing level, “and then a long plateau above which the student seems incapable of rising.” That is the intermediate plateau, described in 1899.

The finding has been contested. A 1958 paper by F. S. Keller in the Journal of the Experimental Analysis of Behavior carries the title “The phantom plateau”. Only its bibliographic record could be opened for this article, so its argument is not summarised here. Whether a plateau is a real stage of learning or an artefact of how the curve was drawn is still the right question, and the modern data answers “both, depending on the plateau”.

What does the research actually show about plateaus and learning curves?

Table 1 sets out the studies opened for this article, with figures as printed in each source.

Table 1. The research record on plateaus and learning curves (verified figures)

Source Data What it reported
Bryan and Harter (1899), Psychological Review Individuals tested weekly; consensus of about two hundred operators Receiving curve shows plateaus followed by sudden ascents; sending curve does not
Rosenbloom and Newell (1982), AAAI proceedings Summary of the power law of practice “performance on a task improves as a power-law function of the number of times the task has been performed”
Heathcote, Brown and Mewhort (2000), abstract only 40 data sets, 7,910 learning series, 475 subjects, 24 experiments Exponential function fit better than the power function in all the unaveraged data sets; averaging biased results toward the power law
Donner and Hardy (2015), Psychonomic Bulletin and Review 25,280 individual learning curves, 500 measurements each, four cognitive tasks A piecewise model beat a single smooth curve; later pieces surpassed earlier ones “after a brief drop in performance at the transition point”
Gaschler and colleagues (2014), Frontiers in Psychology 1,383 chess players with at least 10 years of tournament play Exponential curve fit the first 5 years better for 88% of players and predicted later years better for 62%; the third of players with the lowest early gains showed small rating decreases in the first years
Gray and Lindstedt (2017), abstract only Reanalysis of two classic studies plus a third data set Dips mark exploration of new methods; leaps mark a successful new method
Ericsson (2008), Academic Emergency Medicine Review Everyday skills are automated after “frequently less than 50 hours”; after that, more experience does not improve accuracy
Choudhry, Fletcher and Soumerai (2005), abstract only 62 evaluations of physicians, MEDLINE 1966 to June 2004 32 (52%) found decreasing performance with years in practice for all outcomes; 13 (21%) for some outcomes; 13 (21%) no association; 2 (3%) concave; 1 (2%) increasing for some outcomes; 1 (2%) increasing for all
Stafford and Dewar (2014), Psychological Science, accepted version 854,064 players of an online game; 45,672 played more than 9 times Players who spread their first ten plays over more than 24 hours later scored 47,264 on average against 44,050 for those who did not (Cohen’s d = 0.11)
Soderstrom and Bjork (2015), abstract only Review of motor and verbal learning Performance during practice “is often an unreliable index” of long-term learning

Sources: as named in each row; full references in the Sources section. Rows marked “abstract only” rest on the published abstract, not the full paper.

Four things stand out.

First, the smooth curve is a property of averages. Heathcote and colleagues, and later Donner and Hardy, found that the tidy power law appears when many learners are pooled. An individual’s line has flats and steps in it.

Second, the steps have a shape. Donner and Hardy describe a gradual improvement with diminishing gains inside one strategy, and a discrete sequence of strategy shifts, each better in the long run than the last. Gray and Lindstedt give the pattern its name: plateaus, dips and leaps.

Third, experience is not practice. Summing the physician review’s categories, 45 of 62 evaluations (72.6%) found performance declining with years in practice for at least some outcomes, and 2 of 62 (3.2%) found it rising for at least some. The review reports associations, not causes, and is cited here as research on skill maintenance, not as guidance about choosing a clinician.

Fourth, what a learner sees during practice is not what the learner has learned. Soderstrom and Bjork separate performance, which can be observed in the session, from learning, which only shows later. A flat week can hide real learning.

How much of a plateau is simply arithmetic?

More than most learners assume. Two calculations make the point.

The first uses Cambridge’s own estimates. A Cambridge University Press guide to the Common European Framework of Reference (CEFR) for teachers lists the guided learning hours learners typically need, and adds: “You will notice that it takes longer to progress a level as learners move up the scale.”

Table 2. Guided learning hours by CEFR level, with the cost of each step (CEOtudent analysis of Cambridge University Press figures)

CEFR level Guided learning hours as printed Midpoint Step from the level below (midpoint) Weeks for the step at 5 guided hours a week (illustrative assumption)
A1 Approximately 90-100 95 95 19
A2 Approximately 180-200 190 95 19
B1 Approximately 350-400 375 185 37
B2 Approximately 500-600 550 175 35
C1 Approximately 700-800 750 200 40
C2 Approximately 1,000-1,200 1,100 350 70

Source for column 2: Cambridge University Press (2013), Introductory Guide to the Common European Framework of Reference (CEFR) for English Language Teachers. Columns 3-5 are CEOtudent calculations that read the printed figures as cumulative totals from zero; the 5 hours a week is an illustrative assumption, not a Cambridge figure.

Read this way, the step into B1 costs 1.95 times the step into A2, and the step into C2 costs 3.68 times as much. There is a doubling inside the table: zero to B1 takes about 375 hours at the midpoints, and B1 to C1 takes another 375. Zero to B2 is about 550, and B2 to C2 is another 550. A learner on a steady 5 hours a week collects a new label every 19 weeks at the start and waits 70 weeks for the last one. Nothing has gone wrong. The milestones have moved further apart.

A caution: the guide’s wording is that these are hours “to progress between levels”. If each figure is the cost of that level alone, the escalation is steeper, with C2 at 11.6 times A1. The direction is the same under either reading.

The second calculation uses the power law of practice as described by Rosenbloom and Newell: time on a task falls as a power function of the number of repetitions. The three exponent values below are illustrative assumptions, chosen only to show the shape.

Table 3. Where the improvement arrives under a power law of practice (CEOtudent analysis; exponent values are illustrative assumptions)

Exponent (illustrative assumption) Time per attempt at repetition 100, as % of the first attempt Share of the total gain between repetition 1 and 10,000 that has arrived by repetition 10 By repetition 100 By repetition 1,000 Share left for repetitions 1,001-10,000 Extra practice needed for the next 10% cut in time
0.2 39.8% 43.9% 71.5% 89.0% 11.0% +69% of all practice so far
0.3 25.1% 53.2% 79.9% 93.3% 6.7% +42% of all practice so far
0.4 15.8% 61.7% 86.3% 96.1% 3.9% +30% of all practice so far

Source: CEOtudent arithmetic on the power-law form described by Rosenbloom and Newell (1982); reproduced in calc.py. No measured data set is behind these rows.

The last column is the plateau in one number. With an exponent of 0.3, a further 10% cut in time always requires 42% more practice than the learner has done in total. At repetition 100 that is 42 more repetitions. At repetition 1,000 it is 421. At repetition 10,000 it is 4,208.

The exponential curve that fits individuals better tells the same story. Gaschler and colleagues give a worked example: a chess player who starts at 1,000 rating points, will end at 1,500, and gains about 20% of the remaining gap each year gains 100 points in year one and 80 in year two. Extending their example with the same rule gives 64 points in year three and 13.4 in year ten. By the end of year three the player has covered 48.8% of the whole distance.

This is the arithmetic plateau: what a healthy learning curve looks like from the inside. The fix is to measure over longer windows and to decide whether the next step is worth its price. For the hours side of that question, see time-to-competence data for 20 in-demand skills.

Why does more experience stop producing improvement?

The second cause is automaticity. Ericsson’s 2008 overview, drawing on the stage account of skill acquisition by Fitts and Posner, describes the sequence. A newcomer makes obvious errors, works out adequate responses, and with further experience produces them with less and less effort. Then: “When performance has reached this level of automaticity and effortless execution, additional experience will not improve the accuracy of behavior”.

Ericsson’s phrase for what follows is arrested development. Automation is the goal of the beginner phase and the trap of the intermediate one. Once typing, driving, a tennis serve, a code-review habit or a standard paragraph structure runs without attention, it also runs without correction. The learner has lost what the overview calls the ability “to make specific intentional adjustments”.

The same paper lists the conditions under which practice has been uniformly associated with improvement: a task with a well-defined goal, motivation to improve, feedback, and ample opportunities for repetition and gradual refinement. Many intermediate learners have quietly lost two of the four: the goal has become “do the activity”, and feedback has thinned out because nothing fails visibly any more. How far deliberate practice explains skill differences across domains is a separate debate, covered in the 10,000-hour piece.

This is the reverse of Bryan and Harter’s plateau. Theirs came from too little automation of the basics; Ericsson’s comes from too much. The two feel identical and need opposite treatment, which is why diagnosis comes before effort.

Why does performance often dip before it leaps?

Because a better method is, at first, a worse-performing one. Gray and Lindstedt describe dips as “the marker of exploration where performance suffers as new methods are invented and tested”, and leaps as the moment a successful new method is put to work. Donner and Hardy saw the same thing in 25,280 curves and asked, without answering, why transitions are usually accompanied by a temporary drop.

A self-taught typist who switches to touch typing gets slower before getting faster. A programmer who starts writing tests first ships less for a while. A chess player who abandons a comfortable opening loses games. A writer who starts outlining produces fewer words. The old method had been polished for years and the new one is on day one.

The dip should therefore be planned for: time-boxed, and measured against the trend of the old method, not its best day.

Which plateau are you on? A six-type diagnosis

Table 4 separates the mechanisms. It is a CEOtudent editorial framework: the six types and their names are ours, and each rests on the sources named in the row.

Table 4. The CEOtudent Plateau Diagnosis (editorial framework, built on the sources in Table 1)

Plateau type Diagnostic sign Research it rests on Intervention
1. Arithmetic plateau Progress is still visible over a quarter or a year, but not week to week Power law of practice; Gaschler’s exponential curve; Cambridge hours per level Lengthen the measurement window; price the next step; make the stop, maintain or push decision
2. Measurement plateau The yardstick is one coarse number (a level, a rating, a general feeling) and sub-skills have never been tested separately Soderstrom and Bjork on learning versus performance; Richards on learners who “do not perceive” progress Break the skill into 4-6 sub-skills, test each, and retest after a delay instead of at the end of a session
3. Bottleneck plateau One basic component still needs conscious attention and everything above it queues behind it Bryan and Harter’s hierarchy of habits; Richards on limited vocabulary range Isolate the component and drill it until it runs without attention
4. Automaticity plateau The activity feels easy, errors are rare and small, and nobody has corrected the learner in months Ericsson’s arrested development; Choudhry’s physician review Raise the standard, slow down, restore a specific goal and outside feedback
5. Method ceiling Scores cluster tightly near a personal best, and better performers are visibly doing something different Gray and Lindstedt; Donner and Hardy Change the method and accept a planned dip
6. Load and spacing plateau Volume is high, sessions are massed, and quality falls within a session Ericsson’s daily limit on deliberate practice; Stafford and Dewar on spaced play Cut volume, space sessions, protect recovery

Two of these deserve a note.

On the measurement plateau: Jack Richards, writing on the move from intermediate to advanced English, lists five features of the plateaued learner, among them a gap between receptive and productive competence, persistent fossilised errors and fluency that has progressed at the expense of complexity. A learner improving in listening while frozen in speaking is hidden by a single “level”.

The load plateau has a published reference point. Ericsson’s handbook chapter reports that elite performers’ practice “never consistently exceeds five hours per day” and describes a limit of four to five hours of daily deliberate practice across domains. In Stafford and Dewar’s data, spreading the first ten plays over more than 24 hours was worth about 3,000 points, which the authors equate to an extra 50% of practice at that stage. The gap in means was 3,214 points, or 7.3%, a small effect on a game task. The wider spacing literature is covered in spaced repetition beyond flashcards. This is a summary of research findings and general information only, not training or health advice.

The decision test

Answer in order and stop at the first “yes”.

  1. Compared with the same test three months ago, not last week, is there measurable improvement? If yes, arithmetic plateau.
  2. Is progress judged by one overall number or by feel? If yes, measurement plateau until sub-skill tests say otherwise.
  3. Under time pressure, does one basic component (vocabulary recall, syntax, fingering, calculation) still demand conscious attention? If yes, bottleneck.
  4. Has it been more than a month since anyone or anything gave specific corrective feedback, and does a typical session feel comfortable? If yes, automaticity.
  5. Are the last ten results tightly grouped near a personal best while stronger performers use a visibly different technique or tool? If yes, method ceiling.
  6. Is practice long, daily and massed, with quality dropping inside the session? If yes, load and spacing.

If nothing fits, return to question 2. An unmeasured skill cannot be diagnosed.

Should this plateau be broken at all?

This is the CEO question, and it comes before any protocol. A plateau is the point where the price of the next unit of skill becomes visible. An owner looks at that price and asks what it buys.

Table 5. Stop, maintain or push (CEOtudent editorial framework)

Decision Choose it when What it looks like
Stop The skill no longer serves a current goal, or a tool or colleague now covers it well enough Take it off the learning list, record the level reached, release the hours
Maintain The current level already clears the bar the work requires, and the next level would not change outcomes A small spaced maintenance dose; no improvement target; no guilt
Push The next level changes what can be earned, built or decided, and the diagnosis points to a fixable cause The 4-week protocol below, with one plateau type targeted

Maintaining is a legitimate strategy, and so is stopping, the subject of how to choose what not to learn. What the CEO lens rules out is the most common option: the same routine, at the same hours, with a bad feeling about the result.

How do you break through? A 4-week protocol

This is the student side: change the design of practice, not the hours. The durations are editorial suggestions, not research findings.

Table 6. The 4-week breakthrough protocol (CEOtudent editorial framework)

Week Job What to do Evidence it borrows from
1. Measure Replace the single yardstick Split the skill into 4-6 sub-skills. Record a baseline for each under realistic conditions. Run the decision test. Pick one plateau type and one sub-skill Soderstrom and Bjork; Richards
2. Redesign Change one thing about the practice Bottleneck: isolated drills. Automaticity: a harder standard plus outside feedback. Method ceiling: adopt the new method and write down the expected dip. Load: halve session length and space sessions Ericsson; Gray and Lindstedt; Stafford and Dewar
3. Hold Stay in the dip Keep the new design even though scores fall. Compare against the pre-change trend, not the personal best. Keep feedback specific and immediate Donner and Hardy; Ericsson
4. Retest Check learning, not performance After at least two days without practising that sub-skill, repeat the week-1 test. Then decide: continue, switch plateau type, or move to maintain Soderstrom and Bjork

Worked across domains, the redesign week looks like this. Languages: stop adding input and record five minutes of speech for correction of fossilised errors. Coding: rebuild a small tool without an assistant, then compare with a reviewed reference. Chess: analyse lost games before playing new ones. Music: play the hard bars at half tempo with a recording. Sport: film the movement and change one cue. Writing: get line edits on one page instead of producing ten more.

Does AI cause plateaus or help break them?

Both. The reasoning in this section is editorial; the studies behind it are discussed in the linked pieces.

AI assistance can produce an automaticity plateau faster than any earlier tool. When a model drafts the paragraph, completes the function or supplies the translation, output becomes fluent while the learner’s own retrieval and error correction are never exercised. Two of Ericsson’s four conditions, repetition and feedback on one’s own attempt, disappear. The finished work also stops saying anything about the unaided skill. The pattern is examined in deliberate practice in the age of AI and in the project-based learning ladder.

The same tools are well suited to three of the interventions. They can give specific, immediate feedback on an attempt the learner made alone, which answers feedback starvation. They can generate sub-skill tests on demand, which answers the measurement problem. They can produce endless isolated drills for a bottleneck component. The rule that keeps the benefit and avoids the cost is simple: attempt first, then ask. A tutor posture for this is laid out in the 20-hour AI tutor protocol.

One test settles it. Once a month, do a representative task with no assistance and compare it with the same task from three months earlier. If assisted output keeps rising while unaided output is flat, the plateau is real, and the tool is hiding it.

What this does not mean

  • It does not mean every plateau can be broken. The studies describe patterns and mechanisms; none promises a leap after a dip.
  • It does not mean the tables predict an individual’s curve. Table 3 is arithmetic on assumed exponents, and the evidence in Table 1 is that individual curves depart from any smooth function. In the chess data, a substantial share of players fit neither curve.
  • It does not mean the Cambridge hours measure a given learner. The guide calls them estimates and notes that learners vary; the cumulative reading is ours.
  • It does not mean the evidence is uniform across skills. Much of it comes from telegraphy, online games, cognitive training tasks and chess ratings, and four of the ten rows in Table 1 rest on abstracts only.
  • It does not mean that slow early starters are finished. In the chess data, players with no early gains caught up to some extent later, although they did not reach the level of fast starters by year 10.

Frequently asked questions

Why does progress slow down after the beginner phase?
Learning curves flatten by design: under a power law with an illustrative exponent of 0.3, 79.9% of the improvement available across 10,000 repetitions arrives in the first 100. Early skills also become automatic and stop being corrected.

How long does an intermediate plateau last?
No source opened for this article gives a general duration. In Cambridge’s estimates the B1 and B2 steps each cost roughly twice the A2 step, so at constant weekly hours the intermediate levels take about twice as long.

Is a plateau a sign of reaching a personal limit?
Not necessarily. The sources here describe automaticity, coarse measurement and method ceilings as causes, and Ericsson’s account treats the stall as arrested development that changed practice can restart.

Should practice hours be increased to break a plateau?
Not as a first move. Experience alone showed no reliable benefit in the physician review, and Ericsson reports a limit of four to five hours of daily deliberate practice. Change the design of the hours first.

Why did performance get worse after changing technique?
That is the expected dip. Donner and Hardy found a brief drop in performance at the transition between strategies, with later pieces of the curve typically surpassing earlier ones.

Which study techniques help most at the intermediate stage?
The ones that restore retrieval and feedback. The ranking in what the evidence says about learning covers them in order of research strength.

Sources

  1. Bryan, W. L., and Harter, N. (1899). Studies on the telegraphic language: The acquisition of a hierarchy of habits. Psychological Review, 6(4), 345-375. Text of the scanned volume held by the Internet Archive.
  2. Keller, F. S. (1958). The phantom plateau. Journal of the Experimental Analysis of Behavior, 1, 1-13. (bibliographic record only; full text not opened, argument not summarised)
  3. Rosenbloom, P. S., and Newell, A. (1982). Learning by chunking: Summary of a task and a model. Proceedings of the AAAI-82 National Conference on Artificial Intelligence.
  4. Heathcote, A., Brown, S., and Mewhort, D. J. K. (2000). The power law repealed: The case for an exponential law of practice. Psychonomic Bulletin and Review, 7(2), 185-207. (abstract only)
  5. Donner, Y., and Hardy, J. L. (2015). Piecewise power laws in individual learning curves. Psychonomic Bulletin and Review, 22(5), article starting at page 1308. Open-access full text.
  6. Gaschler, R., Progscha, J., Smallbone, K., Ram, N., and Bilalic, M. (2014). Playing off the curve: Testing quantitative predictions of skill acquisition theories in development of chess performance. Frontiers in Psychology, 5, 923. Open-access full text.
  7. Gray, W. D., and Lindstedt, J. K. (2017). Plateaus, dips, and leaps: Where to look for inventions and discoveries during skilled performance. Cognitive Science, 41(7), 1838-1870. (abstract only)
  8. Ericsson, K. A. (2008). Deliberate practice and acquisition of expert performance: A general overview. Academic Emergency Medicine, 15(11), 988-994.
  9. Ericsson, K. A. (2006). The influence of experience and deliberate practice on the development of superior expert performance. In The Cambridge Handbook of Expertise and Expert Performance, chapter 38. Cambridge University Press.
  10. Choudhry, N. K., Fletcher, R. H., and Soumerai, S. B. (2005). Systematic review: The relationship between clinical experience and quality of health care. Annals of Internal Medicine, 142(4), 260-273. (abstract only)
  11. Stafford, T., and Dewar, M. (2014). Tracing the trajectory of skill learning with a very large sample of online game players. Psychological Science, 25(2), 511-518. Accepted version, White Rose Research Online.
  12. Soderstrom, N. C., and Bjork, R. A. (2015). Learning versus performance: An integrative review. Perspectives on Psychological Science, 10(2), 176-199. (abstract only)
  13. Richards, J. C. (2008). Moving Beyond the Plateau: From Intermediate to Advanced Levels in Language Learning. Cambridge University Press booklet, author’s copy.
  14. Cambridge University Press (2013). Introductory Guide to the Common European Framework of Reference (CEFR) for English Language Teachers.

Note on the tables. Table 1 reports figures as printed in the sources. Tables 2 and 3 are CEOtudent analyses: Table 2 recomputes Cambridge’s printed guided learning hours into per-step costs under a cumulative reading, and Table 3 is arithmetic on the power-law form with illustrative exponents. The year-three and year-ten chess figures extend a worked example printed by Gaschler and colleagues. Tables 4, 5 and 6 and the decision test are CEOtudent editorial frameworks. Deliberately not used: the popular “OK plateau” label and any summary of Keller’s 1958 argument, because no primary text was opened; the Foreign Service Institute language-hour categories and the deliberate-practice variance percentages, which sibling articles cover; and a secondary description saying the exponential function won in 33 of 40 data sets, because the authors’ own abstract states the result differently.


This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.

This post is also available in: Türkçe Français Español Deutsch

Benzer içerikler