Don’t read the manual. Let the manual come to you, daily.

Ashisu: a source-locked language tutor, trained to withhold the answer.

Research stage. Nothing is trained yet. Open source on completion.

Abstract

Ashisu is a language-learning companion that teaches only from a source the learner approves, cites every answer, and is trained, not prompted, to withhold the answer and guide the learner up a nine-level hint-ladder. It is at the research stage. We hold 25,966 supervised examples, 9,794 preference pairs, an evaluation pipeline and one measured baseline: before any fine-tuning, the base model handed over the answer before the learner had tried twice in 69.6% of tutor turns (78 of 112). Phase 1 will fine-tune the model and measure six things separately. We are looking for funding, compute credits, and advisors or reviewers. This page sets out the mechanism, the baseline, the published evidence and its limits, the formulas we build on, and the plan.

Request the research brief

1 Introduction

Ashisu is a goal-driven, source-locked language-learning companion. The learner names a goal, Ashisu finds a source for it, and the learner approves that source. From then on everything taught comes from that source, with citations, in five to ten minutes a day.

You name the goal. Write it in your own words: “Pass the IELTS”, “Reach B2 in English”, “Speak English at my new job”. If the goal is already in the catalogue, you can start straight away.
You approve the source. Ashisu finds candidates, such as the official syllabus or course book with its edition and date. The one you approve is the only source it teaches from.
Five to ten minutes a day. A 140-page course book becomes a daily rhythm: a micro-lesson, a question and spaced review.
When the source changes, you learn the difference. A new edition is not taught from scratch. Starting from what you already know, Ashisu teaches only what changed.

“It is not in this source.”The one reply when the source has no answer. We would rather say nothing than invent something.

Ashisu does not claim to know the truth. It teaches the source you approved, and every answer carries a citation back to it. Ashisu is for language learning only. The product direction is every language; the current research and model cover English as a second language, and the project rejected general multilingual instruction data because it teaches direct-answer leakage.

2 The tutor

The source lock decides what Ashisu may teach. The tutor decides how. It is meant to be trained, not prompted, to keep the answer back and to guide the learner up a nine-level hint-ladder, starting at the lowest rung that could work and climbing only as far as the learner needs. The full answer is the last resort. That discipline is to be trained through supervised fine-tuning and preference optimisation, and whether it holds up in long conversations is what Phase 1 tests.

“…I use the tracks ___ run.”

Tutor: “What is the purpose here? Try: I use the tracks ___ run.”

Figure 1. A real exchange from the training data (first language Turkish, English level B1). The tutor points to the spot and leaves the word for the learner. It is an example from the data, not output of a trained model.

**Table 1.** The hint-ladder (R1–R9). Figure 1 is an R2 example.
Rung	What the tutor does
R1	Encourage: “try again”
R2	Locate the error and leave it blank
R3	Name the rule, with a different example
R4	Narrow the choice
R5–R8	In the research brief
R9	Give the word, then have the learner rewrite the line

The tutor adapts to the learner’s level on the CEFR scale; see Section 5.6.

3 Phase 1 measurements and baseline

Phase 1 measures six things separately, on full multi-turn dialogues of at least ten turns, and runs the same dialogues through the base model, a prompt-only version, SFT and, if needed, DPO. Table 2 shows where each measurement stands.

**Table 2.** The six Phase 1 measurements.
Measurement	What it asks	Base model (baseline)	After fine-tuning
Answer leakage (early disclosure)	Share of tutor turns that hand over the answer before the learner has tried twice. Lower is better.	69.6% (78 of 112 turns)	Not measured: nothing is trained yet
Diagnosis accuracy	Does the tutor identify correctly what the learner got wrong?	In progress	Not measured
Hint graduation	Does help start low and climb only as far as the learner needs?	In progress	Not measured
Over-refusal	Does the tutor hold back help the learner needs? A model can look safe simply by refusing to teach.	In progress	Not measured
Learner’s own words	Did the learner finish the work in their own words? Judged on a blinded multi-turn test.	Not yet measured	Not measured
Teacher usefulness	How useful the tutor’s behaviour is judged to be from a teacher’s point of view.	Not yet measured	Not measured

3.1 The baseline

The one reading we have was taken on the base model, Qwen3-30B, before any fine-tuning: 16 held-out scenarios of 7 turns each, judged by Claude Sonnet 5. The model handed over the answer before the learner had tried twice in 69.6% of tutor turns (78 of 112), and in 66 of those 78 it did so before the learner had tried at all. Giving the answer after the learner has insisted is legitimate teaching, and it is measured separately: 0% on the baseline. The fine-tune has to beat 69.6% without increasing over-refusal.

The baseline was measured on the earlier Qwen3-30B pick. The current direction is Qwen3.8-27B.

3.2 Known gaps

Multi-turn behaviour data is thin: 1,214 dialogues. Whether that is enough is unknown until the run.
About 2–3% of feedback records still contain word-level leakage.
There is no human evaluation yet, only an LLM judge.

4 Published evidence

These are the closest studies we have found. None of them is an Ashisu result. Our own model has not been trained yet, and Phase 1 is the next step. Our review covered 44 sources, 18+ papers and 7 training recipes.

**Table 3.** Published studies, what they measured and where they stop. All results are reported by others.
Study	Setting	Result	Where it stops
Bastani et al. (PNAS 2025)	Randomised classroom study at a Turkish high school: nearly 1,000 students, four 90-minute maths sessions, GPT-4.	Unrestricted GPT-4 raised practice scores by 48% but lowered the later closed-book exam by 17%. A version prompted to give hints, with teacher-written solutions, removed the harm and produced no gain.	Mathematics, one school, a 2023 model.
Dinucu-Jianu et al. (EMNLP 2025)	Online reinforcement learning with a leakage penalty, against offline fine-tuning and multi-turn DPO.	Dialogues that gave away the solution fell from 89.5% to 5.4%, at some cost to the simulated student’s progress. Offline SFT (36.0%) and multi-turn DPO (35.6%) did not go below the untuned model (29.3%).	Mathematics, a simulated student, LLM judges, one seed.
PEARL (arXiv 2605.29582)	A Socratic-tutor recipe trained from Qwen3-30B-A3B, a 30-billion-parameter mixture-of-experts model.	Its leakage-control score rose from 41.1 to 76.5 and its average evaluator score from 69.1 to 84.2. In a blind comparison by human teachers it was preferred to its base model in 64.1% of cases.	The authors’ own benchmarks and scale.
TACT (arXiv 2608.03952)	A 4-billion-parameter model trained with reinforcement learning, tested on 78 authentic English-tutoring contexts.	Answer-leaking or over-helping fell from 0.346 to 0.026, and the overall score rose from 0.629 to 0.832.	The same LLM judge supplied the training reward and scored the test. Preprint.

4.1 What the evidence does not settle

Real learners. Only one of these studies measured learning in real students, and it was in mathematics. The others rely on simulated students or LLM judges, and two of them report no check of those judges against human raters.
Offline or online training. One study found online reinforcement learning far more effective than offline fine-tuning or multi-turn DPO. Another reported gains from fine-tuning followed by DPO, but measured them with a simulated student (Scarlatos et al. 2025). Which recipe suits a language tutor is open.
Model size. The English-tutoring result comes from a 4-billion-parameter model. PEARL used a 30-billion-parameter model, the class closest to Ashisu’s Qwen3.8-27B.
Turkish-speaking learners. The published learner data we found cover English learners whose first language is Spanish (Settles et al. 2018) and Korean-speaking learners of English (Kwon et al. 2024). None covers Turkish speakers learning English, and the Turkish high-school study was in mathematics.
CEFR data. The CEFR-labelled reading corpus in one readability study had 331 passages and no A1 texts (Xia, Kochmar & Briscoe 2016).
Hint-ladders. We found no study that compares a graded hint-ladder with simply withholding the answer. The English-tutoring corpus we checked has no graded hints, and by our reading of its figures, its direct-answer act is far more frequent than its self-correction act (Kwon et al. 2024).

5 Formulas and studies we build on

Ashisu does not invent its own rules for practice, difficulty or feedback. It builds on published models, listed here with what each has shown and where it stops. Figures are given as the authors reported them, and we are still cross-checking each one.

Effect sizes are differences in standard deviations: d = (M₁ − M₂) / SD_pooled, and Hedges’ g = d · (1 − 3 / (4·df − 1)) corrects d for small samples. Values near 0.2 are small, 0.5 medium and 0.8 large.

5.1 When to practise again

Memory fades with time, and each successful recall slows the fading. Three published models describe this, and a scheduler can use any of them to decide when an item is due.

Half-life regression (Settles & Meeder 2016)
  p = 2^(−Δ/h)              recall probability p after a gap Δ, for half-life h
  h = 2^(Θ·x)               half-life predicted from practice history x (seen, recalled, missed)

FSRS-6 forgetting curve (community specification, not peer-reviewed)
  R(t, S) = (1 + f·t/S)^(−w),   f = 0.9^(−1/w) − 1
  S: stability (R = 0.9 when t = S), w: trained decay

ACT-R with spacing (Pavlik & Anderson 2005)
  m = ln Σ t_i^(−d_i),   d_i = c·e^(m_(i−1)) + α,   p(recall) = 1 / (1 + e^((τ − m)/s))
  m: activation, t_i: time since practice i, d_i: its decay, τ: threshold, s: noise

Spacing works for languages. A meta-analysis of 48 second-language experiments (98 effect sizes, N = 3,411) found spaced practice better than massed practice, with a medium-to-large effect. Longer gaps helped on delayed tests but not on immediate ones, and expanding gaps did no better than equal ones (Kim & Webb 2022; confirmed from the abstract only).
Testing beats restudying. Across 159 effect sizes from 61 studies, being tested gave better later recall than studying again (Hedges’ g = 0.50). The gain was larger when the answer was shown afterwards (g = 0.73 against 0.39) and when the final test came a day or more later (0.69 against 0.41). These were laboratory studies, mostly with undergraduates, and none gives a second-language estimate (Rowland 2014).
Promising, but not proven on learning. On 12.9 million Duolingo word traces, half-life regression predicted recall with a mean absolute error of 0.128 (the Leitner system: 0.235), but its ranking accuracy was near chance (AUC 0.538), and the live experiments measured engagement, not learning (Settles & Meeder 2016). In a 30-day shared task, spacing features added 0.013 AUC, which was not significant (Settles et al. 2018).
Parameters are not universal. The fitted decay parameters of the ACT-R model varied several-fold across datasets, and its data are laboratory recall of single words by adults (Pavlik & Anderson 2005).

5.2 Knowing the learner, choosing the next question

To pick the next question, the tutor needs an estimate of what the learner knows and of how hard each item is. Three models provide that estimate, and a fourth result suggests how hard the next question should be.

Rasch model (item response theory)
  P(correct) = 1 / (1 + e^(δ − θ))            θ: learner ability, δ: item difficulty

Elo rating for learning (Pelánek 2016)
  θ ← θ + K·(correct − P)        δ ← δ + K·(P − correct)
  K = a / (1 + b·n),  a = 1, b = 0.05          n: answers so far

Bayesian knowledge tracing (Corbett & Anderson 1995)
  P(correct) = L·(1 − S) + (1 − L)·G           L: skill learned, S: slip, G: guess
  L_n = L′ + (1 − L′)·T                         L′: L updated on the answer, T: learning per attempt

The 85% rule (Wilson et al. 2019)
  ER = F(−β·Δ)             Δ: difficulty, β: learner precision, F: noise distribution
  ER* = F(−1) ≈ 0.159      fastest learning: about 85% accuracy, for Gaussian noise

CEFR can serve as a difficulty scale. The Duolingo English Test puts items and test-takers on a 100-point scale aligned with CEFR (A1 = 0, A2 = 20, B1 = 40, B2 = 60, C1 = 80, C2 = 100) and serves items close to the test-taker’s ability. Its scores correlate at r = 0.74 with TOEFL iBT and r = 0.75 with IELTS, in a study by its own developers using self-reported scores. Machine-estimated word difficulty agreed only moderately with expert labels on unseen words (r = 0.56) (Settles, LaFlair & Hagiwara 2020).
No model wins everywhere. On nine tutoring datasets, logistic regression led on four and deep knowledge tracing on five. On the only language dataset (Spanish, 182 learners), logistic regression reached an AUC of 0.863 against 0.832 for deep knowledge tracing, and simple models tended to win below about a million interactions (Journal of Educational Data Mining). Extended Bayesian knowledge tracing matched deep knowledge tracing on average (Khajah, Lindsey & Mozer 2016).
Elo is cheap but needs data. It needs roughly 100 learners for stable item difficulties. In simulation it recovered difficulty from adaptively chosen questions where raw success rates failed (Pelánek 2016).
Mastery gating helped in one study. Practising until P(learned) reached 0.95 raised the share of students scoring 90% or more from 24% to 56% in a programming tutor. The study does not say how students were assigned to conditions (Corbett & Anderson 1995).
The 85% rule is a simulation result. It comes from neural-network simulations without human participants, assumes binary classification and Gaussian noise (other noise laws give 82% or 75%), and contains no forgetting term (Wilson et al. 2019).

5.3 Matching text to the learner’s level

A tutor that teaches from a source must also judge how hard that source is for the learner, and rewrite it when it is too hard.

Vocabulary coverage
  coverage = known running words / all running words × 100

Zipf–Mandelbrot law (Piantadosi 2014)
  f(r) ∝ 1 / (r + β)^α          f: frequency of the word with rank r
  fitted to American English: α = 1.13, β = 2.73, R² = 0.91

CEFR control error (Malik et al. 2024)
  ControlError(x, t) = (s_cefr(x) − t)²        t: target level (A1 = 1 … C2 = 6), s_cefr: CEFR scorer

Lexical diversity (definitions as implemented in open-source code)
  HD-D = (1/n) · Σ over word types [ 1 − C(N − f, n) / C(N, n) ],   n = 42
  MTLD = N / (F + (1 − TTR_r) / (1 − θ)),   θ = 0.72,   averaged over a forward and a backward pass
  N: tokens, f: occurrences of a type, F: completed factors, TTR_r: type–token ratio of the remainder

Coverage thresholds are soft. About 95% known words for listening and 98% for reading are the figures most often cited, with about 90% for audiovisual material. Only six studies measured coverage precisely, mostly with advanced learners and few texts, and none used CEFR labels. On a 400-word page, 98% coverage means about 8 unknown words (Webb 2021).
Vocabulary size by level varies by language. Recognising the following numbers of the 5,000 most frequent English lemmas has been linked to CEFR bands: A2 1,500–2,500; B2 3,250–3,750; C2 4,500–5,000. Mean scores differed between English, French and Greek learners, and no Turkish learners were sampled (Milton 2010).
Readability tools must be tested on learner texts. A classifier trained on texts for native readers reached only 0.233 accuracy on CEFR-graded passages. A ranking model transferred far better (pairwise accuracy 0.913), and mapping its score to CEFR bands gave about 62% exact-level accuracy (Xia, Kochmar & Briscoe 2016).
Small models need training to hold a level. Prompting alone gave GPT-4 a control error of 0.28–0.57, but 7B models stayed between 1.19 and 2.76. Fine-tuning followed by reinforcement learning brought a 7B model to 0.39, level with GPT-4 given a detailed prompt. The same scorer served as reward and as metric, which risks rewarding the scorer rather than the level (Malik et al. 2024).
Diversity indices disagree. MTLD was reported to be independent of text length (McCarthy & Jarvis 2010; abstract only), but a later study of learner essays found HD-D the index least affected by length (Bestgen 2024, author manuscript). HD-D with a fixed sample size is also deterministic, which keeps evaluation runs repeatable.

5.4 What kind of help works

The hint-ladder rests on two findings and one open question.

Feedback that explains beats bare right or wrong. Across 435 studies (994 effect sizes), feedback had a mean effect of d = 0.48. Feedback that explained what was wrong and how to fix it reached d = 0.99 (42 effect sizes), right-or-wrong plus the correct answer d = 0.46, and praise or punishment d = 0.24. Seventeen per cent of all effect sizes were negative. The review has no language-learning breakdown and does not separate withholding the answer from giving it (Wisniewski, Zierer & Hattie 2020).
Prompting the learner beats reformulating for them. In a meta-analysis of 15 classroom studies (N = 827), oral corrective feedback had significant, lasting effects, and prompts that push the learner to self-correct did better than recasts that supply the correct form. Effects were clearest on free-production tests and larger for younger learners (Lyster & Saito 2010; abstract only).
How much help is right remains open. Koedinger and Aleven (2007) call this the assistance dilemma and say the best balance is unresolved. The tutoring systems they review combine step-by-step feedback, messages for common errors and on-demand hints, and the support for each feature varies. The R1–R9 ladder is our answer to the dilemma, not a settled result.

5.5 Rewarding the right behaviour

Published training recipes reward a tutor when the student makes progress and penalise it when it gives the answer away. Three of them are written out here.

Dinucu-Jianu et al. 2025 (online reinforcement learning)
  r = r_solve + λ·(r_pedagogy − 1)
  r_solve = (1/K) · Σ 1[ŝ_k = s]     share of K sampled answers that are right after the dialogue
  r_pedagogy = Π 1[judge accepts]    1 only if every judge accepts the dialogue (no leak, good tone)

TACT, Yang et al. (preprint)
  r = r_task + α·r_ground + β·r_format + γ·r_strategy + (δ·a − λ·l − μ·o)
  a: acceptable, l: leaks or over-helps, o: off-task;   α = 0.2, β = 0.1, γ = 0.3, δ = 0.2, λ = μ = 1

Preference score for DPO pairs (Scarlatos et al. 2025)
  s = λ·y + (1 − λ)·r,   λ = 0.5
  y: predicted chance that the student’s next reply is correct, r: rubric score

Proxies, not learning. All three rewards rely on a simulated student or an LLM judge, so each measures a stand-in for learning. Results and limits of these recipes are in the next section.

5.6 CEFR levels

The Common European Framework of Reference (CEFR) is the international standard for describing language ability. Schools use it, and exams such as IELTS and TOEFL report scores that map onto it. The language the tutor uses, the errors it targets and how far up the ladder it climbs all adapt to the learner’s level.

A1Beginner

A2Elementary

B1Intermediate

B2Upper intermediate

C1Advanced

C2Mastery

6 Set against the alternatives

Generic assistants and most prompt-wrapped “AI tutor” apps are built to give a good answer quickly, which is exactly the behaviour a tutor has to resist. Ashisu is designed to teach, and Phase 1 will test whether it does.

Ashisu compared with generic assistants and prompt-wrapped tutor apps
Capability	Generic assistantChatGPT, Gemini	Prompt-wrapped app	Ashisudesign target
Withholds the answer	answers at once	fragile, relies on the prompt	trained into the model
Hint-ladder scaffolding	none	ad hoc	R1–R9 taxonomy
Holds up over long dialogues	reverts to giving answers	scaffolding collapses	trained against collapse (DPO)
Adapts to the learner’s CEFR level	generic	superficial	aligned with A1–C2
Teaching quality is measured	not measured	not measured	two-way, judge-scored metric
You own the model and data	closed API	rents an API	Apache-2.0 base model, self-hosted

by design, to be confirmed in Phase 1 partial or prompt-dependent absent

Published tests support the first row. Untuned or prompted general models gave away the full solution in 29–61% of tutoring dialogues (Dinucu-Jianu et al. 2025), and strong general models gave away the answer or over-helped in 9–27% of authentic English-tutoring contexts (TACT, preprint).

7 Plan and what we ask for

The next step is Phase 1: fine-tune on the curated data and measure the result against the baseline. Training runs on a rented 80 GB-class GPU at a modest cost per run.

7.1 Phase 1 checklist

Held-out set: 16 scenarios (done)
Baseline measured: 69.6% early disclosure on the base model (done)
Training data curated and verified: 25,966 SFT examples, 9,794 preference pairs (done)
Evaluation pipeline built (done)
SFT with LoRA on a rented 80 GB GPU (open)
Measure after SFT and compare with the baseline (open)
DPO, if SFT still leaks (open)
Final comparison: base, SFT, DPO and SFT then DPO, with a guard against reward hacking (open)
A small human evaluation, if possible (open)

7.2 What we ask for

Funding. Grants for the training runs and their evaluation. Amounts are in the research brief.
Compute credits. GPU hours or cloud credit for Phase 1, on an 80 GB-class GPU.
Advisors and reviewers. Experts who will criticise the method and the metric, and help design an evaluation with real learners.

7.3 What supporters receive

The research brief now, with the full hint-ladder and the amounts. Everything the research produces is meant to be shared openly, and Ashisu will be released as open source when it is complete.

Plan registration: not yet registered. The Phase 1 plan will be registered publicly, and the record will be linked here once it exists.

8 Technical summary and status

8.1 Under the hood

Base model: Qwen3.8-27B, a dense 27B model under Apache-2.0 (released August 2026). This is the current direction, pending a fine-tuning trial and a VRAM re-check.
Method: Supervised fine-tuning (SFT), then preference optimisation (DPO)
Training data: 25,966 teaching examples and 9,794 preference pairs (curated and verified)
Evaluation: An LLM-judge pipeline with a two-way metric (answer leakage and over-refusal), scored record by record
Why 27B: Large enough to apply the full nine-rule rulebook (in our own tests, models under 4B could not follow it reliably)

8.2 Status log

Model selected and validated
Training data curated and verified
Evaluation pipeline built
Teaching policy and metric defined
Literature review of formulas and evidence (cross-checking in progress)
Phase 1: training and evaluation (next)
Open-source release on completion

9 Frequently asked questions

What can I learn with Ashisu?

Languages, and only languages. Ashisu is built for language learning, and the research begins with English as a second language.

What if the source doesn’t answer my question?

Ashisu replies: “It is not in this source.” It would rather say nothing than invent an answer.

How long does a day of study take?

Five to ten minutes: a micro-lesson, a question and spaced review.

Why short sessions spread over several days?

Because spaced practice beats massed practice. A meta-analysis of 48 second-language experiments found a medium-to-large advantage, and longer gaps paid off when retention was tested later. Expanding gaps did no better than equal ones, so the schedule does not need to be elaborate (Kim & Webb 2022).

What happens when the source is updated?

You are not taught it from scratch. Starting from what you already know, Ashisu teaches only what changed.

Does the tutor just give me the answer?

No. It is trained to withhold the answer, point to the error and guide you up a hint-ladder. The full answer is the last resort.

Is withholding the answer proven to help?

Giving answers freely is shown to harm: in a randomised study at a Turkish high school, students using unrestricted GPT-4 did worse on the later exam (Bastani et al. 2025). A tutor that gave hints instead removed the harm but did not, on its own, improve results. We found no published study that tests a graded hint-ladder on language learners, and whether it improves language learning is the question this research is built around.

Will Ashisu be open source?

Yes. Ashisu will be released as open source when it is complete. The base model is already Apache-2.0, and every artefact of the research is meant to be shared openly.

When can I use it?

Ashisu is still in development, and no release date has been announced.

Whom do I contact about my account, privacy or data?

Write to support@ashisu.com. The same address handles account questions and privacy or data requests.

10 Request the research brief

The research brief sets out the method, the metric, the plan and the amounts in full. If you fund or support work like this, through grants, compute or as an advisor or reviewer, write and ask for it.

Request the research brief

For account help, privacy or data requests, or any other question, write to support@ashisu.com. This page sets no cookies and uses no analytics.

References

Bastani et al. (2025). Proceedings of the National Academy of Sciences 122(26), e2422633122. doi:10.1073/pnas.2422633122
Bestgen (2024). Study of lexical-diversity indices on learner essays (author manuscript).
Corbett & Anderson (1995). User Modeling and User-Adapted Interaction 4(4), 253–278. doi:10.1007/BF01099821
Dinucu-Jianu et al. (2025). Proceedings of EMNLP 2025, 272–292. doi:10.18653/v1/2025.emnlp-main.15
FSRS community specification (2026). “The Algorithm”, open-spaced-repetition project wiki. Not peer-reviewed.
Journal of Educational Data Mining, article 451. Comparison of knowledge-tracing models on nine datasets. Link
Khajah, Lindsey & Mozer (2016). Proceedings of Educational Data Mining 2016.
Kim & Webb (2022). Language Learning 72(1), 269–319.
Koedinger & Aleven (2007). Educational Psychology Review 19(3), 239–264. doi:10.1007/s10648-007-9049-0
Kwon et al. (2024). Proceedings of ACL 2024, 3389–3414. doi:10.18653/v1/2024.acl-long.186
Lyster & Saito (2010). Studies in Second Language Acquisition 32(2), 265–302. doi:10.1017/S0272263109990520
Malik et al. (2024). Findings of ACL 2024. doi:10.18653/v1/2024.findings-acl.926
McCarthy & Jarvis (2010). Behavior Research Methods 42(2), 381–392. doi:10.3758/BRM.42.2.381
Milton (2010). EUROSLA Monographs 1, 211–232. PDF
Pavlik & Anderson (2005). Cognitive Science 29(4), 559–586. doi:10.1207/s15516709cog0000_14
Pelánek (2016). Computers & Education. doi:10.1016/j.compedu.2016.03.017
Piantadosi (2014). Review of Zipf’s law in natural language (author manuscript). PDF
Rowland (2014). Psychological Bulletin 140(6), 1432–1463. doi:10.1037/a0037559
Scarlatos et al. (2025). AIED 2025, LNCS 15877. doi:10.1007/978-3-031-98414-3_18
Settles & Meeder (2016). Proceedings of ACL 2016, 1848–1858. doi:10.18653/v1/P16-1174
Settles et al. (2018). Proceedings of the BEA workshop, 56–65. doi:10.18653/v1/W18-0506
Settles, LaFlair & Hagiwara (2020). Transactions of the ACL 8, 247–263. doi:10.1162/tacl_a_00310
Webb (2021). Reading in a Foreign Language 33(2), 287–302.
Wilson, Shenhav, Straccia & Cohen (2019). Nature Communications 10, 4646. doi:10.1038/s41467-019-12552-4
Wisniewski, Zierer & Hattie (2020). Frontiers in Psychology 10, 3087. doi:10.3389/fpsyg.2019.03087
Xia, Kochmar & Briscoe (2016). Proceedings of the BEA workshop, 12–22. aclanthology.org/W16-0502
Yang et al. (preprint). TACT. arXiv:2608.03952