Writing study: Mastering Sparse CUDA Generation through Pretrained Models and Deep Reinforcement Learning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Yaoyu Wang, Hankun Dai, Zhidong Yang, Junmin Xiao, Guangming Tan · 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: Learning & optimization algorithms (primary) · Inference, algorithms & systems efficiency.

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

38 physical pages

Before references / appendix

≈ 10.2 page-equivalents

Main figures / tables

5 / 3

Attached appendix figures / tables

3 / 17

Attached appendix span

≈ 22.65 page-equivalents

Before references / appendix: ≈ 26.84%References / intervening material: ≈ 13.55%Attached appendix: ≈ 59.61%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	≈ 1.82 pp.	4.79%
2 RELATED WORK	3	≈ 1.13 pp.	2.97%
3 SPARSERL	4	≈ 2.53 pp.	6.66%
4 EXPERIMENT	7	≈ 1.73 pp.	4.55%
5 ABLATION STUDY	8	≈ 1.78 pp.	4.68%
6 CONCLUSION	10	≈ 0.5 pp.	1.32%
ACKNOWLEDGMENTS	11	≈ 0.09 pp.	0.24%
REFERENCES	11	≈ 5.15 pp.	13.55%
A APPENDIX	16	≈ 22.65 pp.	59.61%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 4
• Table 1 · p. 8
• Figure 3 · p. 8
• Figure 4 · p. 9
• Table 2 · p. 9
• Table 3 · p. 9
• Figure 5 · p. 10

Attached appendix

• Table 4 · p. 18
• Table 5 · p. 18
• Figure 6 · p. 20
• Table 6 · p. 20
• Table 7 · p. 21
• Table 8 · p. 22
• Table 9 · p. 23
• Table 10 · p. 23
• Table 11 · p. 24
• Table 12 · p. 24
• Table 13 · p. 26
• Table 14 · p. 26
• Table 15 · p. 27
• Table 16 · p. 28
• Table 17 · p. 29
• Table 18 · p. 29
• Table 19 · p. 30
• Figure 7 · p. 32
• Figure 8 · p. 34
• Table 20 · p. 38

Argument and organization

1. Introduction and related work

   Define sparse kernel generation as input-structure-dependent optimization rather than ordinary functional code generation.

2. SparseRL

   Stage the pretrained model, structural matrix interface, supervised initialization, and hardware-feedback RL loop.

3. Experiments, ablations, and discussion

   Separate functional correctness from execution efficiency, then inspect component contributions and reuse costs.

4. Appendix A

   Collect diagnostics, implementation artifacts, additional comparisons, costs, failures, and reviewer-facing clarifications.

Problem, claim and contribution

The opening two-kernel example grounds the problem in different correct implementations with different runtimes. This motivates a hardware-aware objective more concretely than a generic claim about code-generation capability.

Section 1; Figure 1, page 2. · Main paper

Argument and section order

The ablation sequence follows the architecture: training stages, structural embeddings, and RL choices. Matched-backbone and budget comparisons are useful but placed late in the appendix, making fairness harder to assess from the main argument.

Section 5, pages 8–10; Appendix A.24, page 33. · Main paper

Figures, captions and visual explanation

Figure 2 aligns three training stages in separate panels and distinguishes initialization, forward execution, and red dashed backpropagation. Its caption explains the role of the actor and critic rather than merely naming boxes.

Figure 2, page 4. · Main paper

Tables, comparisons and evidence

Table 1 reports model size, pass@1/pass@5/pass@1000, and compilation rate separately; Figure 3 bins performance by nonzero count and uses boxplots rather than hiding matrix heterogeneity behind one mean.

Table 1 and Figure 3, page 8. · Main paper

Appendix and reproducibility

Concrete code examples, shared-memory measurement, prompt templates, failure matrices, and generation-versus-execution costs make the appendix operationally useful. However, future-tense proposed comparisons are mixed with completed experiments.

Appendices A.10, A.15–A.17, A.25–A.29, pages 25–38. · Main paper

Clarity and writing conventions

Some appendix headings and paragraphs retain response-to-reviewer language. A final archival paper should convert those into self-contained claims and clearly distinguish preliminary studies from established results.

Appendices A.26–A.28, pages 33–35. · Main paper

What to reuse—and what not to infer

Transferable lesson: Use a concrete task-level example, a stage-aligned training diagram, and separate correctness/runtime evidence; surface matched controls and deployment costs early enough to support the central claim.

Boundary: Do not inherit all numerical prose: Section 4.2 assigns the Qwen2.5-14B 49.25/57.50 results to CodeT5, whose Table 1 values are 48.75/56.50; Appendix A.9 says greater exploration reduces correctness although its table increases pass@k; dataset counts and speedup summaries also differ across appendices. Sinusoidal-frequency interpretation in A.2 reverses the small-/large-index scale relation.

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 1–6; acknowledgments on page 11. Reference list excluded from analytical reading.

Appendix / supplement

Pages 16–38, Appendix A.1–A.29, including code listings, reward details, controlled baselines, failure cases, cost analyses, and prompts.

Visually inspected PDF pages

4, 8

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Wed Feb 11 09:34:21 2026 EST · 38 pages