Three Pillars of Language Learning
Three Pillars of Language Learning
Language study time splits into three modes. Each session gets one of the three, chosen by how much of the session’s content the learner already understands. Two of the modes are immersion. Immersion means time spent with material made for native speakers, with the learner’s attention on it. Material made for learners does not count. Audio playing in another room while attention is on something else does not count. Immersion does not mean living abroad. The third mode is preparation, and it is the only one of the three that happens away from real content.
Preparation is the only mode that produces no exposure to the language. The other two modes use identical material at different tool settings. The standard freeflow move shows this: content already worked through with subtitles and lookups gets rerun later with both off. So each session carries two decisions. The first is how much time goes to study away from content. The second is where the tool dial sits on the content itself, and a count of unknown words decides that one.
Core takeaways
- Each session gets one of three modes: preparation, interactive immersion, or freeflow immersion. A count of unknown words in a hundred-word stretch decides the tool setting, and the count takes about half a minute.
- Zero to two unknowns per hundred words means tools off. Around five means subtitles and a small lookup budget. Ten or more means different content, or groundwork before the content reopens.
- Priming, meaning a small set of words, sounds, characters or one grammar pattern loaded before input, pays over months. In a single evening it often leaves comprehension unchanged, and knowing the topic first plus a second listen did more.
- Freeflow supplies the hours. Extensive reading, across 34 studies and 3,942 participants, gave d = 0.71 from before to after and d = 0.46 against controls.
- Preparation runs under a time cap. The next change goes to whatever is limiting sessions now: comprehension, attention, or hours. This bottleneck rule is a house rule with no research behind it.
- No mode in the model produces a spoken or written sentence. The start date for production has to be set from outside the model.
Counting unknown words
The tool setting comes from a count of the words that do not resolve, and the count takes about half a minute. Take one dense stretch of roughly a hundred words, a screen of subtitles or a paragraph of a novel, and read it through with a tally of every word that fails to come through.
Reading research supplies the two edges of the scale. Comfortable unassisted comprehension sits near 98% known words. Around 95% is the floor where comprehension still holds up with support. Below that floor the aids stop helping. Subtitles, a popup dictionary, slowed playback and a plot known in advance all raise how much of a piece is understood, and none of them rescues material with ten unknowns in every hundred words.
| Unknown words per hundred | What the session is | What changes |
|---|---|---|
| 0–2 | followed at full speed | tools come off, the session runs freeflow |
| around 5 | followable with support | subtitles, a small lookup budget, one more pass |
| 10 or more | under the comprehension floor | different content, or groundwork before this content reopens |
A paragraph of a novel, a hundred words counted off, gives eleven that do not resolve. Eleven is under the floor, so working through that paragraph is a lookup exercise and does not count as a session. A chapter by the same author, already read in translation, gives four unknowns per hundred. Four routes to interactive: audiobook subtitles on, four lookups, one pass.
The count reports on one stretch of one piece of content, so it gets rerun whenever the content changes. A denser chapter of the same book can land two rows lower in the table than the chapter before it.
What priming does
Priming means loading a small set of words, sounds, characters or a single grammar pattern shortly before input, aimed at recognition rather than recall. Its payoff over hundreds of hours is not in dispute. The 95% and 98% coverage figures describe that payoff: a larger known vocabulary means more of any given page is understood. Over one evening the effect is small and unreliable. Pre-teaching a word list before a listening session reliably raises knowledge of those words and raises learner confidence, and it often leaves comprehension of that session unchanged. In a direct comparison, knowing the topic of the passage beforehand plus a second listen beat the vocabulary support.
That result changes the order of the levers for a session that is going badly. The content lever is the cheaper one. A plot summary read first, or a rewatch of something already known in the first language, moves comprehension further than an hour spent building a deck for the content.
Preparation is the mode that feels most like progress while it is happening, so its time is fixed by a cap and the learner does not judge it session by session. The field’s standard allocation splits study four ways, roughly a quarter each: language-focused study, meaning-focused input, meaning-focused output and fluency work. The daily shape this model inherited is smaller and stricter: ten minutes on the writing system, fifteen on vocabulary priming, twenty inside the language, forty-five minutes in total, with priming capped so any spare time goes to input. Either anchor lets yesterday’s time be checked against a number.
Where the hours come from
Freeflow is the mode that supplies volume. A session that stops at every unknown word contains fewer minutes of language than one that runs straight through. The measured version of this comes from extensive reading, the same low-friction shape applied to text. Across 34 studies and 3,942 participants, extensive reading gave d = 0.71 from before to after and d = 0.46 against controls, across reading speed, comprehension, vocabulary and grammar together. The model calls what accumulates over that volume instinct: meaning arrives with no translation step in between.
Background audio and attention
The failure named against freeflow is content playing while attention is elsewhere, logged afterwards as immersion. As an absolute rule this overshoots. Attention multiplies the value of exposure, and exposure with no attention on it still counts for something. The claim that acquisition requires attention has been contested for decades. Adults who heard non-native tone contrasts for eight hours across four days, with their attention elsewhere the whole time, afterwards showed faster and larger neural responses to those contrasts. Sound-level familiarity is close to all that unattended exposure produces. That is enough to give background audio a small slot labelled phonology, and it stays out of the immersion total.
Choosing what to change next
Balancing by bottleneck is a working rule made in this vault, and no research supports it. Whichever single thing is limiting sessions now gets the next change: comprehension, attention, or the raw hours going in. No matching construct was found in the field, so its whole support is the general case for adjusting a method while the method is still running.
The rule also disagrees with the quarter-each allocation. Under the bottleneck rule, a week of preparation and nothing else is allowed when low comprehension is stopping every session. Under fixed proportions it is not allowed. The reconciliation used in the model lets the bottleneck choose what changes next, under a cap that keeps the exception from becoming the schedule.
When attention is the bottleneck, the first move is a change of content, before any harder technique: a shorter session, or content already seen once that carries its own visual support. That ordering is untested. Its nearest evidence is the listening comparison in which knowing the topic beforehand plus a second listen beat the vocabulary support.
The old worry about looking things up too often is another unproven house rule. No study locates a lookup density at which comprehension breaks down. The nearest evidence runs the other way. Glossed reading is lookup with the friction removed, and it produces roughly 45% word learning against 27% for unglossed reading when tested immediately afterwards, while comprehension of the text improves. The rule that survives is about how a session ends: the lookup budget worth setting is the one that keeps the session running to its last line.
Where the model comes from and what it costs
The model descends from comprehensible-input pedagogy. That pedagogy’s central formulation, content pitched one step above the current level, has never been given an operational definition. That absence is the standing reason for calling it untestable. A recent critique goes further and argues that the absorption framing sits badly against evidence that acquisition is active, interactive and multimodal. The model does not depend on that hypothesis. The measured results are what support it.
The model fails the reader who has to hold a conversation in three months, because every hour it routes goes to input. The price of using it is a decision before each session, the half-minute count, and a record of where the hours actually went. Without that record the balancing rule cannot run. The larger cost is that labelling sessions becomes work in itself. An hour of subtitled video with four lookups is interactive and freeflow at once, and forcing it into one box can cost more than the routing gives back.
Those costs are worth paying only while the routing changes something. If two weeks of counting leaves every session looking the same, the model is no longer changing anything, and the time goes back to plain input. The check that can fail is re-entry. Content worked through interactively should be followable at full speed within a session or two of that first pass. A second pass that still needs the tools means the first pass was tool operation rather than comprehension.
Nothing in the model produces a sentence
The model is input only, and nothing in it says so. No mode in it produces a spoken or written sentence, while the standard four-way allocation gives meaning-focused output a full quarter of study time. The omission is inherited rather than chosen: the source material these three modes were compressed from puts production off to a later stage it never describes. A year can pass under this model with nothing produced, and no part of the model reports a problem. A date set on the calendar is the only check.
The three modes reduce to one routing rule with a number at each edge. An evening’s content gets counted, and the count names the tool setting for that evening. Study time away from content stays under a ceiling, so it cannot take the hours the content needs. Production is the one thing the count never routes to, so its start date has to be set from outside the model.
How to practice this
- Take one dense stretch of about a hundred words, a screen of subtitles or a paragraph of a novel. Count every word that does not come through. Notice that the count takes about half a minute.
- If the count is zero to two, turn the tools off and run the session freeflow. Notice whether the content is followed at full speed.
- If the count is around five, turn subtitles on, set a small lookup budget, and plan one more pass. Notice whether the budget keeps the session running to its last line.
- If the count is ten or more, pick different content, or do groundwork before this content reopens. Notice that subtitles, a popup dictionary, slowed playback and a known plot do not rescue content this dense. Recount whenever the content changes, since a denser chapter of the same book can land two rows lower.
- Keep preparation under a time cap. The inherited daily shape is ten minutes on the writing system, fifteen on vocabulary priming, twenty inside the language. Notice whether yesterday’s minutes can be checked against that number.
- Within a session or two of an interactive pass, rerun the same content with subtitles and lookups off. Notice whether it is followable at full speed. A second pass that still needs the tools means the first pass was tool operation rather than comprehension.
Related pages
- Preparation: the forms this mode takes, from sound and script work to tool setup, with the failure table for the week when preparation starts taking the input hours.
- Interactive Immersion: the activity list behind the tools-on setting, and the conditions that make a lookup-heavy session worth running at all.
- Freeflow Immersion: the entry conditions for tools-off work, plus the short post-session note on comprehension, interest and focus.
- Hacking Comprehension Menu: the lists of levers for raising comprehension, sorted into tools, techniques and content, with the rule about changing one variable at a time.
- Attention is Important: what attention is for while comprehension is still low. The graded version of the passive-watching rule belongs there.
Open questions
- Does the hundred-word count hold when the unknowns cluster in one technical stretch instead of spreading evenly?
- At what point does labelling a session cost more than the routing saves?
- How early should production enter for a learner whose goal is conversation rather than reading?
- Is a phonology-only passive slot worth keeping once content routinely counts under two unknowns per hundred words?
Sources
Comprehension thresholds: Laufer & Ravenhorst-Kalovski (2010), “Lexical threshold revisited: Lexical text coverage, learners’ vocabulary size and reading comprehension,” Reading in a Foreign Language, https://files.eric.ed.gov/fulltext/EJ887873.pdf · Kremmel et al. (2023), “Unknown Vocabulary Density and Reading Comprehension: Replicating Hu and Nation (2000),” Language Learning, https://onlinelibrary.wiley.com/doi/10.1111/lang.12622 · “How does lexical coverage affect the processing of L2 texts?” Applied Linguistics 45(6), https://academic.oup.com/applij/article/45/6/953/7841943
Balance of study types: Nation (2007), “The Four Strands,” Innovation in Language Learning and Teaching 1(1), open copy: https://www.academia.edu/41764855/The_Four_Strands
Tool-assisted input: Montero Perez et al. (2013), “Captioned video for L2 listening and vocabulary learning: A meta-analysis,” System, https://www.sciencedirect.com/science/article/abs/pii/S0346251X13001012 · Kurokawa (2025), “Incidental Vocabulary Acquisition Through Captioned Viewing: A Meta-Analysis,” Language Learning, https://onlinelibrary.wiley.com/doi/10.1111/lang.12697 · Zhang & Ma (2024), “The effect of textual glosses on L2 vocabulary acquisition: A meta-analysis,” Language Teaching Research, https://journals.sagepub.com/doi/abs/10.1177/13621688211011511 · “The effects of glosses on English L2 incidental vocabulary learning through reading: a meta-analysis,” Frontiers in Language Sciences (2026), https://www.frontiersin.org/journals/language-sciences/articles/10.3389/flang.2026.1815571/full
High-volume unassisted input: Nakanishi (2015), “A Meta-Analysis of Extensive Reading Research,” TESOL Quarterly, https://onlinelibrary.wiley.com/doi/10.1002/tesq.157 · “Learning a Language Through Reading: A Meta-analysis,” Educational Psychology Review (2025), https://link.springer.com/article/10.1007/s10648-025-10068-6
Attention: Schmidt (2010), “Attention, awareness, and individual differences in language learning,” https://nflrc.hawaii.edu/PDFs/SCHMIDT%20Attention,%20awareness,%20and%20individual%20differences.pdf · Truscott (1998), “Noticing in second language acquisition: a critical review,” https://www.sdkrashen.com/content/articles/noticing_1998.pdf · “Dynamic changes in electrical brain activity during passive learning of foreign speech sound discrimination,” PMC, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12262118/
Session-level payoff of priming: Aldukhayel, “The Impact of Vocabulary Preteaching and Content Previewing on L2 Listening Comprehension,” https://www.inform.nu/Articles/Vol26/InfoSciV26p023-038Aldukhayel8865.pdf · Chang & Read, “The impact of vocabulary preparation on L2 listening comprehension, confidence and strategy use,” https://www.researchgate.net/publication/250734037_The_impact_of_vocabulary_preparation_on_L2_listening_comprehension_confidence_and_strategy_use
Lineage and its critics: “Beyond comprehensible input: a neuro-ecological critique of Krashen’s hypothesis in language education,” Frontiers in Psychology (2025), https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2025.1636777/full