Writing study: Self-Programmed Execution for Language-Model Agents

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Luke J O'Connor · 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: Efficiency, inference & systems (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

75 physical pages

Before references / appendix

≈ 9.09 page-equivalents

Main figures / tables

4 / 0

Attached appendix figures / tables

1 / 16

Attached appendix span

≈ 62.91 page-equivalents

Before references / appendix: ≈ 12.12%References / intervening material: ≈ 4%Attached appendix: ≈ 83.88%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	≈ 1.45 pp.	1.93%
2 Formalizing self-programmed execution	3	≈ 1.27 pp.	1.69%
3 A practical language for SPE	4	≈ 1.45 pp.	1.93%
4 Empirical evaluation	5	≈ 2.89 pp.	3.85%
5 Related work	8	≈ 0.84 pp.	1.12%
6 Discussion	9	≈ 0.55 pp.	0.73%
References	10	≈ 3 pp.	4%
A Agentic evaluators and self-programmed execution	13	≈ 10 pp.	13.33%
B Self-programmed execution language for LMs	23	≈ 39 pp.	52%
C Benchmarking methods and results	62	≈ 13.91 pp.	18.55%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 6
• Figure 3 · p. 8
• Figure 4 · p. 8

Attached appendix

• Table 1 · p. 63
• Table 2 · p. 65
• Table 3 · p. 67
• Table 4 · p. 68
• Table 5 · p. 68
• Table 6 · p. 68
• Table 7 · p. 69
• Table 8 · p. 69
• Table 9 · p. 70
• Table 10 · p. 70
• Figure 5 · p. 71
• Table 11 · p. 71
• Table 12 · p. 71
• Table 13 · p. 73
• Table 14 · p. 73
• Table 15 · p. 74
• Table 16 · p. 74

Argument and organization

1. Problem and contribution

   Separates an execution runtime from a fixed orchestration policy, proposing that the model's completion can specify the latter rather than merely select tools within a preset loop.

2. Argument sequence

   A formal definition precedes practical language-design challenges; empirical evaluation distinguishes whether models can use the language, what orchestration they actually choose and how benchmark performance compares.

3. Evidence and verification

   The main text reports weak non-coding results as well as competitive coding results. Source tables retain error cases in denominators, distinguish transport modes and audit unequal spending caps.

Problem, claim and contribution

Separates an execution runtime from a fixed orchestration policy, proposing that the model's completion can specify the latter rather than merely select tools within a preset loop.

Introduction and Figure 1, pp. 1–2 · Main paper

Argument and section order

A formal definition precedes practical language-design challenges; empirical evaluation distinguishes whether models can use the language, what orchestration they actually choose and how benchmark performance compares.

Sections 2–4, pp. 3–8 · Main paper

Figures, captions and visual explanation

Figure 3 plots quality against total estimated run cost in paired benchmark panels and labels reasoning effort at each point; Figure 4 places cost labels on task-success bars across four benchmarks. Figure 1 aligns ReAct, recursive language models and self-programmed execution in three columns. Blue versus yellow identifies logic supplied by the harness versus written by the model, while dashed runtime boundaries remain in all three. The caption explicitly prevents the mistaken inference that model-written orchestration removes the external runtime.

Figures 3–4, p. 8 (render inspected); Figure 1, p. 2 (additional rendered inspection) · Main paper

Tables, comparisons and evidence

The main text reports weak non-coding results as well as competitive coding results. Source tables retain error cases in denominators, distinguish transport modes and audit unequal spending caps.

Section 4, pp. 7–8; Appendix C.4–C.6, pp. 67–71 · Main paper

Appendix and reproducibility

The unusually long appendix is partly executable-language documentation rather than extra experiments; separate source tables, pricing assumptions and trace statistics support each main cost/accuracy plot.

Appendix B, pp. 23–61; Appendix C, pp. 62–75 · Main paper

Clarity and writing conventions

The single-author first-person voice is precise about what was implemented and tested. The discussion separates potential expressivity from capabilities models do not spontaneously use.

Sections 4 and 6, pp. 5–9; Appendix C.8, p. 73 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a new agent runtime, distinguish formal possibility, implementation semantics and observed behavior, and publish source values and cost conventions for each headline plot.

Boundary: The seventy-five-page preprint is not a conference submission-length template. API costs are dated, and different harness prompts, transports and budget caps can affect comparisons.

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 scan across manuscript framing, section sequence, figure/table captions, selected argument/evaluation prose and conclusion; not exhaustive technical or proof verification. Sections 1–6, pp. 1–9; reference boundary pp. 10–12. The scan distinguished formal expressivity from observed agent behavior. Source is the exact public author version listed below, not a verified accepted/camera-ready revision.

Appendix / supplement

Writing-oriented appendix/documentation scan: Appendices A–C, pp. 13–75: abstract-machine theory, language/runtime semantics and full prompts/documentation, then benchmark configurations, costs and trace analyses. Source code, data, model outputs and external video assets were not executed or independently validated.

Visually inspected PDF pages

2, 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 2605.06898v1; not verified identical to the accepted or camera-ready OpenReview revision · 75 pages