Writing study: LittleLearner: Language Models Under Pedagogically-Controlled Knowledge Exposure

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Fanfei Li, Jana Zeller, Manuel Prada-Corral, Thaddäus Wiedemer, Prasanna Mayilvahanan, Ryan Cotterell, Wieland Brendel · Official NeurIPS 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Datasets, benchmarks & measurement (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

32 physical pages

Before references / appendix

≈ 10.09 page-equivalents

Main figures / tables

9 / 1

Attached appendix figures / tables

8 / 6

Attached appendix span

≈ 17.45 page-equivalents

Before references / appendix: ≈ 31.53%References / intervening material: ≈ 13.94%Attached appendix: ≈ 54.53%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Source-checked 2026-10-02: replace two numbered contribution bullets with real Sections 2-3 on page 3. Remove a reference-title false Appendix A on page 13. On page 15 the first A-E lines are a contents list; the real Appendix A starts lower on that same physical page, at position 14.5484. Keep the later B-E section starts.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1 Introduction	2	≈ 1 pp.	3.13%
2 Related Work	3	≈ 0.48 pp.	1.5%
3 Constructing the LITTLELEARNER-LITTLECURRICULUM Sandbox	3	≈ 3.24 pp.	10.13%
4 Probing Capabilities with LITTLELEARNER	6	≈ 2.42 pp.	7.56%
5 Exploration Avenues: Beyond the K–5 Boundary	9	≈ 0.94 pp.	2.94%
6 Conclusion	10	≈ 0.2 pp.	0.63%
Acknowledgements	10	≈ 0.72 pp.	2.25%
References	11	≈ 4.46 pp.	13.94%
A Further Details on the Dataset Filtering Pipeline	15	≈ 8.97 pp.	28.03%
B Details on Model Training	24	≈ 0.31 pp.	0.97%
C Further Evaluation Details	24	≈ 5.59 pp.	17.47%
D More Qualitative Examples	30	≈ 1.68 pp.	5.25%
E Existing Assets, Licenses, and Terms of Use	32	≈ 0.91 pp.	2.84%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Table 1 · p. 2
• Figure 2 · p. 4
• Figure 3 · p. 5
• Figure 4 · p. 5
• Figure 5 · p. 6
• Figure 6 · p. 6
• Figure 7 · p. 7
• Figure 8 · p. 8
• Figure 9 · p. 8

Attached appendix

• Figure 10 · p. 16
• Figure 11 · p. 16
• Table 2 · p. 17
• Figure 12 · p. 17
• Figure 13 · p. 18
• Figure 14 · p. 18
• Table 3 · p. 19
• Table 4 · p. 19
• Figure 15 · p. 26
• Figure 16 · p. 26
• Figure 17 · p. 27
• Table 5 · p. 30
• Table 6 · p. 30
• Table 7 · p. 32

Argument and organization

1. Problem and contribution

   The paper identifies unknown pretraining exposure as a confound in studying knowledge acquisition. Its elementary-school corpus and model are framed as a controlled experimental sandbox rather than a general-purpose language-model replacement or a literal model of child cognition.

2. Argument sequence

   Data filtering and boundary validation precede model capability claims. The experiments then ask whether scaling, post-training and in-context examples change the in-scope versus beyond-scope gap. This sequence makes the validity of the data intervention a prerequisite for interpreting the model results.

3. Evidence and verification

   Matched in-scope and beyond-scope panels separate pretraining and post-training conditions through colors and hatching, with 95% confidence intervals. A second panel set compares prompting interventions; the nearby takeaway specifies the tested 5B model and prompt setups rather than claiming that in-context learning never works.

Problem, claim and contribution

The paper identifies unknown pretraining exposure as a confound in studying knowledge acquisition. Its elementary-school corpus and model are framed as a controlled experimental sandbox rather than a general-purpose language-model replacement or a literal model of child cognition.

Abstract, introduction and Sections 3–4, pp. 1–8 · Main paper

Argument and section order

Data filtering and boundary validation precede model capability claims. The experiments then ask whether scaling, post-training and in-context examples change the in-scope versus beyond-scope gap. This sequence makes the validity of the data intervention a prerequisite for interpreting the model results.

Sections 3–4, pp. 3–8 · Main paper

Figures, captions and visual explanation

Stage-by-stage retention curves show the filter's behavior on both a construction-related benchmark and an external one. The caption states the precision-first trade-off explicitly: very low beyond-scope retention is obtained by discarding substantial in-scope material, not by claiming exhaustive coverage.

Figure 2, p. 4; rendered page inspected · Main paper

Tables, comparisons and evidence

Matched in-scope and beyond-scope panels separate pretraining and post-training conditions through colors and hatching, with 95% confidence intervals. A second panel set compares prompting interventions; the nearby takeaway specifies the tested 5B model and prompt setups rather than claiming that in-context learning never works.

Figures 8–9 and Sections 4.3–4.4, p. 8; rendered page inspected · Main paper

Appendix and reproducibility

The appendix makes the exposure intervention inspectable with filter details, independent leakage checks, removed-text examples and full model-training specifications. Prompt construction, alternative demonstrations, qualitative responses and an asset/license table support reproduction without conflating use with redistribution.

Appendices A–E, pp. 15–32 · Main paper

Clarity and writing conventions

Question-led experimental headings and short, scoped takeaway boxes reduce the burden of reading many intervention results. Stable in-scope/beyond-scope terminology ties the filtering plot to the capability plots; the useful pattern is controlled contrast, not anthropomorphic wording.

Sections 3–4 and conclusion, pp. 3–10 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a controlled-data resource, validate the intervention before interpreting downstream behavior, and reuse the same condition labels across every figure and conclusion.

Boundary: The curriculum boundary is an engineered operational definition, not perfect proof of all training knowledge or a claim about human children. Negative prompting results are limited to the tested models, tasks and demonstrations.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented distributed manuscript scan of Sections 1–6, physical pp. 1–10: section sequence, substantive prose windows, captions, comparison contracts and conclusions. Not exhaustive technical verification.

Appendix / supplement

Writing-oriented distributed scan of Appendices A–E, physical pp. 15–32, including filters, leakage checks, model training, evaluation prompts and asset terms. Not independent proof checking; code and experiments were not executed.

Visually inspected PDF pages

4, 8

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Public author arXiv manuscript 2608.13545v1; not verified identical to the accepted or camera-ready OpenReview revision · 32 pages