Writing study: RefineStat: Efficient Exploration for Probabilistic Program Synthesis

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Madhav Kanda, Shubham Dipak Ugare, Sasa Misailovic · Official ICLR 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: Inference, algorithms & systems efficiency (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

24 physical pages

Before references / appendix

≈ 10.17 page-equivalents

Main figures / tables

2 / 6

Attached appendix figures / tables

0 / 3

Attached appendix span

≈ 9.89 page-equivalents

Before references / appendix: ≈ 42.38%References / intervening material: ≈ 16.42%Attached appendix: ≈ 41.21%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

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	≈ 1.1 pp.	4.58%
2 BACKGROUND	2	≈ 1 pp.	4.17%
3 REFINESTAT	3	≈ 2.85 pp.	11.88%
4 EXPERIMENTAL METHODOLOGY	6	≈ 0.38 pp.	1.58%
5 EXPERIMENTAL RESULTS	6	≈ 3.26 pp.	13.58%
6 RELATED WORK	10	≈ 0.62 pp.	2.58%
7 CONCLUSION AND LIMITATIONS	10	≈ 0.38 pp.	1.58%
8 ACKNOWLEDGMENTS	11	≈ 0.07 pp.	0.29%
REFERENCES	11	≈ 3.93 pp.	16.38%
A APPENDIX	15	≈ 1.33 pp.	5.54%
B PROBABILISTIC PROGRAM METRICS	16	≈ 1.29 pp.	5.38%
C EXAMPLES	17	≈ 1.85 pp.	7.71%
D PROMPT DESIGN	19	≈ 2.25 pp.	9.38%
E EXPERIMENTAL SETUP	21	≈ 0.4 pp.	1.67%
F ERROR ANALYSIS	22	≈ 1.12 pp.	4.67%
G ABLATION STUDY	23	≈ 1.66 pp.	6.92%

All detected figure and table locations

Main text

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

Attached appendix

• Table 7 · p. 18
• Table 8 · p. 23
• Table 9 · p. 24

Argument and organization

1. Executable code versus statistical reliability

   Separates API/grammar validity from posterior diagnostics.

2. Constrained fragments and diagnostic refinement

   Builds a prior–likelihood resampling procedure with an explicit score and search budget.

3. Diagnostic comparisons and failure taxonomy

   Reports execution success separately from inferential quality.

Problem, claim and contribution

A deprecated distribution argument gives a concrete entry point, while the text distinguishes programming semantics from natural-language meaning.

Sections 1–2, pp. 1–3 · Main paper

Argument and section order

Definitions establish each diagnostic before the aggregate reliability score, preventing the score from appearing as an unexplained custom metric.

Sections 2–3, pp. 3–4 · Main paper

Figures, captions and visual explanation

The workflow separates a blue generation block from an orange reliability check; the return arrow makes refinement and its trigger visible.

Figure 1, p. 3 · Main paper

Tables, comparisons and evidence

Paired baseline/method rows expose convergence, effective sample size and divergences beside predictive scores, including explicit failed-program markers.

Tables 2 and 6, pp. 8–9 · Main paper

Appendix and reproducibility

The appendix gives actual prompts, rejected API calls, budget failures and token overhead, which makes the search procedure more reproducible than a success-only demonstration.

Appendices C–G, pp. 17–24 · Main paper

Clarity and writing conventions

The ablation prose describes individually removed components, but the reported delta column uses successive rows; readers cannot safely infer independent contributions from that presentation.

Section 5.4 and Table 4, pp. 8–9 · Main paper

What to reuse—and what not to infer

Transferable lesson: For generated scientific code, report executable validity and scientific diagnostics separately, and show an example that moves through rejection, inference and refinement.

Boundary: Passing five of seven checks is not proof of model adequacy. Predictive checks are omitted, and appendix examples even change the response variable; predictive scores require a common target. Clarify cumulative versus leave-one-out ablations.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–10, Sections 1–7 including diagnostics, constrained generation, search and limitations

Appendix / supplement

Pages 15–24, Appendices A–G: diagnostic definitions, code examples, all prompts, setup, failures and token/memorization analyses

Visually inspected PDF pages

3, 9

• ICLR 2026 final proceedings · 24 pages