Children learn words fast. Everyone knows that. But the shape of that learning curve — how fast, and whether the rate itself changes — has been hiding a structural assumption that a new paper by Michael C. Frank (Stanford) exposes cleanly. And when you apply the same measurement to language models, the curve tells a very different story.
Children, but not language models, show accelerating returns in word learning (arXiv:2608.17120) doesn't propose a new architecture or a new benchmark. It proposes a model class for how children learn vocabulary — and then tests whether LLMs follow the same pattern. They don't. And the difference may explain why children do so much more with so much less data.
The Shape of Learning
The standard view of vocabulary growth is a power law in time: children learn fewer new words per unit time as their vocabulary grows. That's decelerating accumulation. Frank shows this is wrong — or at least, it misses the real story.
The paper fits four candidate growth models to longitudinal vocabulary data from children:
- Constant-rate: same number of new words per unit time
- Decelerating (power law): diminishing returns — fewer new words per unit time as vocabulary grows
- Exponential: proportional growth — each additional unit of experience yields a fixed percentage increase
- Accelerating: each additional unit of linguistic experience yields more learning than the one before
The accelerating model wins. Children don't just accumulate words — they get better at learning as they go. Each new word makes the next one easier. The curve bends upward over developmental time, not downward.
graph LR
subgraph "Children (real)"
A1[Experience 1] -->|"learns 2 words"| A2[Experience 2]
A2 -->|"learns 4 words"| A3[Experience 3]
A3 -->|"learns 8 words"| A4[...]
end
subgraph "LLMs (scaling law)"
B1[Token 1B] -->|"+x% accuracy"| B2[Token 10B]
B2 -->|"+x% accuracy"| B3[Token 100B]
B3 -->|"+x% accuracy"| B4[...]
end
style A1 fill:#27272a,stroke:#22c55e,color:#fff
style A2 fill:#27272a,stroke:#22c55e,color:#fff
style A3 fill:#27272a,stroke:#22c55e,color:#fff
style A4 fill:#27272a,stroke:#22c55e,color:#fff
style B1 fill:#27272a,stroke:#ef4444,color:#fff
style B2 fill:#27272a,stroke:#ef4444,color:#fff
style B3 fill:#27272a,stroke:#ef4444,color:#fff
style B4 fill:#27272a,stroke:#ef4444,color:#fff
What LLMs Do Instead
Frank trains language models on child-directed speech transcripts — the same kind of input a child might hear — and measures their vocabulary learning curves. The result: constant proportional returns. Each doubling of training data yields the same relative improvement. This is exactly what scaling laws predict, and it's the opposite of acceleration.
The gap is stark:
- Children show super-linear returns on linguistic experience. More data doesn't just add more words — it amplifies the learning rate itself.
- LLMs show constant proportional returns. More data helps, but it helps at the same rate per unit data, regardless of how much they already know.
- Children learn using many orders of magnitude less data than language models. The acceleration is a candidate explanation for how.
The finding holds even when LLMs are trained on child-directed speech (rather than web-scale corpora), so it's not just a domain mismatch. The learning dynamics themselves are fundamentally different.
What This Means
This is a small, clean paper — 4 pages, one figure's worth of results — but it hits a nerve. The scaling law story says: more data, same returns per unit data. The child development story says: something qualitative changes as you learn. These are not just different empirical patterns. They suggest different underlying mechanisms.
For people building things:
If the goal is data efficiency — getting more learning per token, not just more tokens — the accelerating model is a design target. Children achieve acceleration because their learning mechanism itself adapts: phonological knowledge bootstraps lexical knowledge, which bootstraps syntactic knowledge, which bootstraps semantic inference. Each layer changes the input processing pipeline.
Current LLMs don't do this. Their architecture is fixed. The transformer stack that processes the first token is the same one that processes the billionth. There's no architectural mechanism for the model to rewire itself to learn differently based on what it already knows.
The paper doesn't propose a solution. But it sharpens the target: if you can build a model whose learning algorithm changes as a function of accumulated knowledge — rather than just its parameters — you might escape the scaling law's constant-returns regime.
Limitations
The comparison is not perfectly controlled. Children and LLMs receive different kinds of input even when both are exposed to child-directed speech — children also get visual, social, and embodied feedback that the models don't. The accelerating model is fit to cross-sectional and longitudinal data, not the result of a controlled intervention. And the paper does not attempt to engineer an accelerating learning mechanism in an LLM — it stops at diagnosing the gap. But diagnosis is where progress starts.
New arXiv submission appearing in cs.CL today, August 19, 2026.
- Children, but not language models, show accelerating returns in word learning — Michael C. Frank, Aug 2026