Writing study: Can In-Context Learning Support Intrinsic Curiosity?

← NeurIPS 2026 oral writing corpus

Eric Elmoznino, Sangnie Bhardwaj, Johannes von Oswald, Rajai Nasser, Blaise Aguera y Arcas, João Sacramento, Rif A. Saurous, Guillaume Lajoie · Official NeurIPS 2026 oral record

Figure, table and section ledger

Editorial contribution type: Theory & formal analysis (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: ≈ 9.66 page-equivalents
Main figures / tables: 4 / 0
Attached appendix figures / tables: 3 / 0
Attached appendix span: ≈ 17.91 page-equivalents

Before references / appendix: ≈ 30.19%References / intervening material: ≈ 13.84%Attached appendix: ≈ 55.97%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Top-level heading and reference/appendix boundaries checked against source PDF text. Bibliography titles/author initials, numbered contribution/property bullets, diagram steps, equations and TOC entries are not substantive headings. Section spans are page-coordinate estimates, not ink area or prose-only length. This is an author manuscript mirror, not a verified accepted-version template. Source layout uses physical single-column y/page-height, including equation-heavy pages.

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	1	≈ 0.78 pp.	2.44%
2 Problem Setting and Notation	2	≈ 0.69 pp.	2.16%
3 Prior Work on Intrinsic Curiosity	3	≈ 0.75 pp.	2.34%
4 Computing Intrinsic Rewards Using In-Context Learners	3	≈ 2.8 pp.	8.75%
5 Experiments	6	≈ 2.96 pp.	9.25%
6 Discussion and Future Work	9	≈ 0.92 pp.	2.88%
Acknowledgments and Disclosure of Funding	10	≈ 0.14 pp.	0.44%
References	10	≈ 4.43 pp.	13.84%
Appendix A Additional Prior Work on Intrinsic Curiosity	15	≈ 0.76 pp.	2.38%
Appendix B Assumptions for Theoretical Results	15	≈ 2.3 pp.	7.19%
Appendix C Proofs for Theoretical Results	18	≈ 8.04 pp.	25.12%
Appendix D Implementing rdl in General BAMDPs vs. BED	26	≈ 0.52 pp.	1.63%
Appendix E Experimental Details	26	≈ 6.28 pp.	19.63%

All detected figure and table locations

Main text

Attached appendix

Argument and organization

Problem and contribution
Poses a feasibility question about prediction-only intrinsic rewards, then separates a negative general case from a constructive restricted case.
Argument sequence
Defines the reward family before its limits, the Bayesian experimental-design exception and three experimental environments.
Evidence and verification
Training curves and final-score comparisons are paired with a spatial noise map and information-gain curves, so reward maximization is not confused with useful exploration.

Problem, claim and contribution

Poses a feasibility question about prediction-only intrinsic rewards, then separates a negative general case from a constructive restricted case.

Introduction; Sections 4.2–4.3 · Main paper

Argument and section order

Defines the reward family before its limits, the Bayesian experimental-design exception and three experimental environments.

Sections 2–5 · Main paper

Figures, captions and visual explanation

Figure 1 distinguishes agent interaction from a pretrained in-context learner; the masked-context inset explains how the reward differs from ordinary surprise.

Figure 1, p. 4 · Main paper

Tables, comparisons and evidence

Training curves and final-score comparisons are paired with a spatial noise map and information-gain curves, so reward maximization is not confused with useful exploration.

Figures 2–3, p. 7 · Main paper

Appendix and reproducibility

Separate mathematical derivations and environment-specific training details let the ideal Bayesian claim and approximate PFN experiments be inspected independently.

Appendices C–E · Main paper

Clarity and writing conventions

The discussion preserves the restricted setting instead of turning positive experimental results into a universal curiosity claim.

Section 6, p. 9 · Main paper

What to reuse—and what not to infer

Transferable lesson: State the negative boundary and the constructive exception before presenting empirical success.

Boundary: BED assumptions and approximate predictive models must remain distinct; this scan does not verify the proofs.

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, pp. 1–9: abstract/framing, section sequence, substantive text windows, captions, evidence organization and conclusion; not exhaustive technical verification.
Appendix / supplement: Writing-oriented scan of Appendix proof and experimental sections, pp. 14–32; section/statement/caption scan, not proof verification; code and experiments were not executed.
Visually inspected PDF pages: 4, 7

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 2606.19476v1; not verified identical to the accepted or camera-ready OpenReview revision · 32 pages