Children do not acquire habits from what their parents tell them. They acquire them from what their parents do and from the environment their parents build. Public-health studies of diet and screen time show this repeatedly, and they show it with a specific method: compare each parent with their own child, then ask which route the influence actually travels. Language acquisition has the same structure, because a child learns morphology (the small meaning-bearing pieces of words, like -s, -ed, un-) from one caregiver's speech. Small language models trained on child-directed speech fail here in a non-human way. They invent irregular forms like bring → brang, where children instead over-regularize to forms like bringed. One suspect sits between the caregiver's speech and the model: the tokenizer, the component that cuts text into units before the model sees it. Standard tokenizers cut words by frequency, not by meaning, so the model may never see the units a child sees.
This note proposes to treat the tokenizer as the learner's "home environment" and to measure its effect with the dyad-level, mediation-style design used in the parenting literature. Each model is trained on one family's caregiver speech, its output is compared with that family's actual child, and the analysis estimates how much of the caregiver-to-learner influence runs through the segmentation. On top of this measurement we test one concrete intervention. The tokenizer learns word boundaries from the caregiver's own partial repetitions ("want the ball? the balls are here"), and is not built from raw frequency. The claim is falsifiable: tokenization causally mediates whether a child-scale model's morphological errors mirror a specific child's errors, and a segmentation built from caregiver repetition structure moves the model toward the child.
What the two directions contribute, and what is open
The behavioral-mirroring literature gives this proposal a method, not a mechanism. Mother–toddler dietary quality is tightly correlated within dyads (Maternal anxiety and diet quality). Child screen time tracks parental screen use rather than stated rules (family factors and screen time). Path analyses find that parental attitudes have no direct effect on adolescent junk-food intake and act only through the home food environment and motivation (Perceived Parental Attitudes…), or through self-efficacy (Modeling Parental Influence…). The transferable ideas are two. Influence should be measured per dyad, not pooled. And the decisive question is which environmental variable carries it. These studies also warn us about their own weaknesses: cross-sectional, self-reported measures with confounded pathways. A simulated learner removes that problem, because we can intervene on the "environment" directly.
The tokenization literature supplies the suspect. Frequency-based tokenizers split words in ways unrelated to morphemes. Byte-pair encoding (BPE, which repeatedly merges the most frequent adjacent character pair) often fuses affixes into stems. Unigram tokenizers isolate negative affixes about 75% of the time. Large models get past bad splits by memorizing, but they fail on nonce words like dis-roagly (Revisiting subword tokenization). Child-scale models cannot memorize, so the cost should be largest exactly where developmental modeling lives.
Developmental modeling supplies the open problem. A word-level acquisition model reached 76% syntactic-category accuracy in English but only 46% in Hebrew, because it treated every inflected string as unrelated. The authors explicitly call for a sub-word layer they did not build (A Language-agnostic Model of Child Language Acquisition). The SIGMORPHON 2022 acquisition task found models learn English -ed from 100 items yet over-irregularize, unlike children. It also found Arabic broken plurals hardest (SIGMORPHON–UniMorph 2022 Task 0), but it never varied subword tokenizers. BabyLM trains on human-scale budgets with off-the-shelf tokenizers (Findings of the Second BabyLM Challenge; BabyLM Turns 4), and also pools all children together.
The obvious bridge, and why we do not take it
The tempting bridge maps "parent teacher, child student" onto cross-tokenizer knowledge distillation. That is training a small model to match a large model's outputs when their vocabularies differ. The obvious "mirroring" variant would have the student learn from teacher-generated data (the environment) instead of teacher token probabilities (the instructions). That variant is essentially sequence-level fine-tuning on teacher outputs, which already exists. It already loses to byte-level distillation: on GSM8K with a Llama-8B teacher and Gemma-2B student, BLD scores 62.55 against 59.29 for fine-tuning and 61.56 for approximate likelihood matching (Cross-Tokenizer LLM Distillation through a Byte-Level Interface; see also Approximate Likelihood Matching). The analogy adds a story but no new prediction, so we drop it. Our proposal differs from distillation in three ways. There is no teacher model. The "parent" is a real caregiver's transcribed speech. The outcome is the match to a real child's errors.
Core Claim
The tokenizer is the dominant environmental mediator between caregiver input and a child-scale model's morphological generalization. Frequency-based BPE trained on child-directed speech produces over-irregularization. It does this by making stem-final strings shared across irregular verbs (the -ing of bring/sing/ring) into reusable units, so the model's analogies run through those strings. A segmentation that places boundaries where the caregiver's own speech alternates should shift errors toward over-regularization and raise the per-family correlation between model and child. Partial repetitions, called variation sets, are the cue: dog/dogs, walk/walked appearing within a few adjacent utterances.
Research Program
The first component is a dyad-matched training setup. We take CHILDES families with both caregiver and child transcripts: English (Adam and other Brown children), German (Leo, also used in SIGMORPHON), and Hebrew (Hagar). For each target family we pretrain a small model on pooled CHILDES speech excluding that family. The pooled data comes from orthographic CHILDES; IPA CHILDES reports 45 million words across 31 languages, which bounds what is available. We then continue training on that family's caregiver utterances in nested, age-ordered slices, following SIGMORPHON's nested-exposure design. At each slice we elicit the model's inflections and compare them with the child's own productions in transcripts at the matched age.
The second component is the intervention, a variation-set tokenizer. Within a sliding window of adjacent caregiver utterances, we find word-type pairs that share a long substring and differ by a short edge string. Those positions become permitted split points. A standard BPE or Unigram tokenizer is then trained with merges forbidden across high-confidence boundaries. Vocabulary size is matched to the baselines. The tokenizer uses only signal a child plausibly has: which forms the caregiver says near each other.
The third component is the mediation analysis, borrowed from the path analyses above. Segmentation quality is the mediator. We measure it as the fraction of target affixes isolated as their own token, using the NegMorph categories of correct, under-segmented and over-segmented. The predictor is caregiver input frequency of each inflectional pattern. The outcomes are model–child agreement and error type. Because we can swap tokenizers while holding input fixed, this is an intervention, not the observational inference the health studies are limited to.
Baselines
The naive reference point is a model that copies the caregiver exactly. It produces each inflected form at the rate the caregiver uses it, and it applies SIGMORPHON's all--ed rule to unseen verbs. If no trained model correlates with the child better than raw caregiver frequency, the whole learning account is empty. Beating it means higher per-dyad correlation with the child's productions than input frequency alone.
The strongest existing morphology learners are the character-level sequence-to-sequence systems from SIGMORPHON 2022 Task 0. They take a lemma plus features and output an inflected form, and they are the published state of acquisition-style inflection modeling. The file gives only qualitative findings for them: they fail to beat the all--ed baseline on irregulars and they over-irregularize. No macro-accuracy number is reported on Lacuna for these systems. We retrain them on identical nested slices with the same tuning budget. Beating them means a higher over-regularization share among errors and higher child correlation at equal inflection accuracy.
The strongest language-model baseline is GPT-BERT, the hybrid causal/masked model that won BabyLM 2024, reporting 86.1 BLiMP and 81.5 SuperGLUE in the 100M-word Strict track (Findings of the Second BabyLM Challenge). We use its architecture as the backbone for every tokenizer condition, and we report BLiMP on English to confirm no condition degrades general grammar.
Tokenizer conditions on that backbone are standard BPE and Unigram at matched vocabulary, plus ByteSpan entropy-based segmentation. We also run byte-level input in the style of CANINE and BanglaByT5. Gold morphological segmentation, in the spirit of linguistically informed splitting (Huck et al.), serves as an upper bound. All conditions see the same words, in the same order, for the same number of epochs. Byte models are compared at both equal words and equal FLOPs, because they process several times more positions.
The key ablation is the variation-set tokenizer trained on the same caregiver utterances in shuffled order. This keeps word frequencies but destroys the adjacency that defines repetitions, mirroring the shuffled control used for CVCL (Grounded language acquisition…). If shuffling does not hurt, the repetition signal is not the ingredient.
Why not just…?
Why not just use byte-level models and skip tokenization? This is a real threat, and SIGMORPHON partly answers it already: its character-level systems still over-irregularize. That suggests removing subwords does not by itself make errors childlike. Byte models also lengthen sequences, which matters under BabyLM-style compute caps. The equal-FLOPs comparison settles it. If byte input matches the variation-set tokenizer on child correlation and error profile at equal compute, we conclude the intervention is unnecessary and report that.
Why not just hyphenate or apply gold morphological splits? Hyphenating affixes lifts Llama-2 negation detection from about 77% to about 96% (Revisiting subword tokenization), so explicit boundaries clearly help. Gold segmentation, however, is not available to a child, so it cannot test whether the learner's environment suffices. It also requires an analyzer per language and does not handle Hebrew root-and-pattern morphology, where meaning sits in non-adjacent consonants (Heinzerling and Strube). We include it as an upper bound. If the variation-set tokenizer recovers most of the gap between BPE and gold, it is the useful finding.
Why not just train on more data? Scale lets large models memorize past bad splits, but the nonce-word failures show memorization is not rule learning (Revisiting subword tokenization). A dyad's caregiver speech is fixed in size by construction, so the question is about learning from one family's input. We nonetheless report the pooled-pretraining size at which BPE's error profile stops changing. This shows whether the tokenizer effect shrinks with data.
Proposed Experiments and Decisive Checks
The primary metric is per-dyad agreement. This is the correlation, across inflectional categories and age slices, between the model's and the child's production rates, the same dyadic statistic used for mother–toddler diet. Secondary metrics are three. One is the error-type ratio on wug-style nonce verbs: over-regularization divided by over-irregularization, where children are above one and SIGMORPHON models were not. Another is NegMorph-style nonce-affix probes. The last is segmentation-quality scores and BLiMP for English.
Four results decide the claim. First, BPE should produce bring-brang-type errors at a higher rate than Unigram or variation-set tokenization. We can check the proposed mechanism directly: those errors should concentrate on verbs whose stem-final string is a shared token. Second, the variation-set tokenizer should beat its shuffled-order ablation on child correlation. If it does not, the intervention fails. Third, the mediation analysis should show that most of the effect of tokenizer choice on child agreement runs through affix isolation. If tokenizers change agreement without changing segmentation quality, the "environment" story is wrong and something else in the tokenizer matters. Fourth, the theory predicts the gain will be smallest in Hebrew, where boundaries are not contiguous. If byte input wins there while the variation-set tokenizer wins in English and German, that typology-dependent pattern is itself the expected and reportable outcome. It would also directly address the 46% Hebrew ceiling that motivated this work.
References
- Maternal anxiety and diet quality among mothers and toddlers from low‐income households
- The impact of family factors and communication on recreational sedentary screen time among primary school-aged children: a cross-sectional study
- Perceived Parental Attitudes Are Indirectly Associated with Consumption of Junk Foods and Sugar-Sweetened Beverages among Chinese Adolescents through Home Food Environment and Autonomous Motivation: A Path Analysis
- Modeling Parental Influence on Food Consumption among Chinese Adolescents through Self-Efficacy: A Path Analysis
- Revisiting subword tokenization: A case study on affixal negation in large language models
- A language-agnostic model of child language acquisition
- SIGMORPHON–UniMorph 2022 Shared Task 0: Modeling Inflection in Language Acquisition
- Findings of the Second BabyLM Challenge: Sample-Efficient Pretraining on Developmentally Plausible Corpora
- BabyLM Turns 4 and Goes Multilingual: Call for Papers for the 2026 BabyLM Workshop
- Cross-Tokenizer LLM Distillation through a Byte-Level Interface
- Universal Cross-Tokenizer Distillation via Approximate Likelihood Matching
- IPA CHILDES & G2P+: Feature-Rich Resources for Cross-Lingual Phonology and Phonemic Language Modeling
- ByteSpan: Information-Driven Subword Tokenisation
- Canine: Pre-training an Efficient Tokenization-Free Encoder for Language Representation
- BanglaByT5: Byte-Level Modelling for Bangla
- LMU Munich's Neural Machine Translation Systems for News Articles and Health Information Texts
- Grounded language acquisition through the eyes and ears of a single child
- Sequence Tagging with Contextual and Non-Contextual Subword Representations: A Multilingual Evaluation