Writing study: OpenThoughts: Data Recipes for Reasoning Models

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Etash Guha, Ryan Marten, Sedrick Keh, Negin Raoof, Georgios Smyrnis, Hritik Bansal, Marianna Nezhurina, Jean Mercat, Trung Vu, Zayne Sprague, Ashima Suvarna, Benjamin Feuer, Leon Liangyu Chen, Zaid Khan, Eric Frankel, Sachin Grover, Caroline Choi, Niklas Muennighoff, Shiye Su, Wanjia Zhao, John Yang, Shreyas Pimpalgaonkar, Kartik sharma, Charlie Ji, Yichuan Deng, Sarah Pratt, Vivek Ramanujan, Jon Saad-Falcon, Stutee Acharya, Jeffrey Li, Achal Dave, Alon Albalak, Kushal Arora, Blake Wulfe, Chinmay Hegde, Greg Durrett, Sewoong Oh, Mohit Bansal, Saadia Gabriel, Aditya Grover, Kai-Wei Chang, Vaishaal Shankar, Aaron Gokaslan, Mike Merrill, Tatsunori Hashimoto, Yejin Choi, Jenia Jitsev, Reinhard Heckel, Maheswaran Sathiamoorthy, Alex Dimakis, Ludwig Schmidt · 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: Datasets, benchmarks & measurement resources (primary) · Empirical analysis, diagnosis & auditing · Learning & optimization algorithms.

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

72 physical pages

Before references / appendix

≈ 9.87 page-equivalents

Main figures / tables

3 / 7

Attached appendix figures / tables

32 / 41

Attached appendix span

≈ 52.89 page-equivalents

Before references / appendix: ≈ 13.71%References / intervening material: ≈ 12.83%Attached appendix: ≈ 73.46%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Independent PDF text and boundary audit: appendix navigation/overview begins before first lettered substantive heading. Physical heading coordinates taken from source-pinned Poppler bbox.

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	≈ 0.86 pp.	1.19%
2 RELATED WORK	3	≈ 0.3 pp.	0.42%
3 OPENTHOUGHTS3 DATA PIPELINE	3	≈ 5.86 pp.	8.14%
4 SCALING OUR PIPELINE TO OPENTHOUGHTS3-1.2M	9	≈ 0.47 pp.	0.65%
5 CONCLUSION	10	≈ 0.18 pp.	0.25%
6 REPRODUCIBILITY STATEMENT	10	≈ 0.09 pp.	0.13%
ACKNOWLEDGEMENTS	10	≈ 0.49 pp.	0.68%
REFERENCES	10	≈ 9.23 pp.	12.82%
Appendix contents	20	≈ 2 pp.	2.78%
A TRAINING DETAILS	22	≈ 1.76 pp.	2.44%
B EVALUATION DETAILS	23	≈ 1.51 pp.	2.1%
C DECONTAMINATION	25	≈ 1.72 pp.	2.39%
D ADDITIONAL SCALING EXPERIMENTS	27	≈ 3.35 pp.	4.65%
E ADDITIONAL DATA RECIPE EXPERIMENTS	30	≈ 5.93 pp.	8.24%
F MODEL REASONING PERFORMANCE ANALYSIS	36	≈ 2.1 pp.	2.92%
G SURPASSING THE TEACHER WITH DISTILLATION FOR LEGAL REASONING	38	≈ 1.21 pp.	1.68%
H ALL TEACHERS ABLATIONS	39	≈ 0.14 pp.	0.19%
I SAFETY ANALYSIS OF OPENTHINKER MODELS	39	≈ 0.76 pp.	1.06%
J EXISTING FRONTIER MODEL EVALUATIONS	40	≈ 0.12 pp.	0.17%
K TESTING REASONING ROBUSTNESS: ALICE IN WONDERLAND EVALUATION	40	≈ 3.4 pp.	4.72%
L COMPUTE REQUIREMENTS	44	≈ 0.15 pp.	0.21%
M SOURCING REASONING TRACES FROM THE WEB	44	≈ 0.85 pp.	1.18%
N LICENSES OF EXISTING ASSETS	45	≈ 1.1 pp.	1.53%
O PIPELINE DETAILS	46	≈ 19.37 pp.	26.9%
P PIPELINE EXPERIMENTS EXPANDED RESULTS	65	≈ 7.42 pp.	10.31%

All detected figure and table locations

Main text

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

Attached appendix

• Table 8 · p. 22
• Table 9 · p. 23
• Table 10 · p. 23
• Table 11 · p. 23
• Table 12 · p. 25
• Figure 4 · p. 26
• Figure 5 · p. 27
• Figure 6 · p. 28
• Figure 7 · p. 29
• Table 13 · p. 30
• Table 14 · p. 30
• Table 15 · p. 31
• Table 16 · p. 31
• Table 17 · p. 31
• Figure 8 · p. 32
• Table 18 · p. 32
• Table 19 · p. 33
• Table 20 · p. 33
• Table 21 · p. 34
• Table 22 · p. 34
• Table 23 · p. 35
• Figure 9 · p. 36
• Figure 10 · p. 37
• Table 24 · p. 37
• Table 25 · p. 37
• Table 26 · p. 38
• Table 27 · p. 39
• Table 28 · p. 39
• Table 29 · p. 40
• Table 30 · p. 41
• Figure 11 · p. 42
• Figure 12 · p. 43
• Figure 13 · p. 44
• Figure 14 · p. 46
• Figure 15 · p. 48
• Figure 16 · p. 49
• Figure 17 · p. 49
• Figure 18 · p. 51
• Figure 19 · p. 51
• Figure 20 · p. 51
• Figure 21 · p. 52
• Figure 22 · p. 52
• Figure 23 · p. 53
• Figure 24 · p. 53
• Figure 25 · p. 56
• Figure 26 · p. 57
• Figure 27 · p. 58
• Figure 28 · p. 60
• Figure 29 · p. 61
• Figure 30 · p. 61
• Figure 31 · p. 63
• Figure 32 · p. 63
• Figure 33 · p. 63
• Figure 34 · p. 63
• Figure 35 · p. 64
• Table 31 · p. 65
• Table 33 · p. 65
• Table 32 · p. 66
• Table 34 · p. 67
• Table 35 · p. 67
• Table 36 · p. 67
• Table 37 · p. 68
• Table 38 · p. 68
• Table 39 · p. 69
• Table 40 · p. 70
• Table 41 · p. 70
• Table 42 · p. 70
• Table 43 · p. 71
• Table 44 · p. 71
• Table 45 · p. 71
• Table 46 · p. 72
• Table 47 · p. 72
• Table 48 · p. 72

Argument and organization

1. OpenThoughts3 data pipeline

   Walks through one empirical design decision at a time, ending each stage with the recipe choice.

2. Scaling and conclusion

   Connects small controlled ablations to a 1.2M-example release and acknowledges untested RL and curriculum choices.

3. Detailed appendices

   Separates implementation, evaluation integrity, extra experiments, asset provenance and full ablation tables.

Problem, claim and contribution

The contribution is a reproducible data recipe rather than a novel model architecture. The introduction names concrete unresolved choices and previews five findings that are then tested stage by stage.

Section 1, pp. 2–3 · Main paper

Argument and section order

Each pipeline subsection follows question, controlled experiment, result and a boxed recipe decision. This gives the long empirical study a navigable argument rather than a catalogue of experiments.

Sections 3.3–3.7, pp. 6–9 · Main paper

Figures, captions and visual explanation

The Sankey-style pipeline exposes dataset attrition, domain mixing and 16-fold answer generation. It conveys quantities and provenance, though pale labels and the extremely short caption reduce standalone legibility.

Figure 3, p. 9 · Main paper

Tables, comparisons and evidence

Table 1 annotates base checkpoints, data availability, optimization type and held-out benchmarks before reporting performance; bolding denotes a two-standard-error band, not just a single winner.

Table 1, p. 2 · Main paper

Appendix and reproducibility

The appendix includes negative results, safety degradation, robustness tests, precise prompts and per-asset licenses. This makes the release auditable beyond headline benchmark gains.

Appendices E, I, K, N–P, pp. 30–72 · Main paper

Clarity and writing conventions

Caption and body checks matter: Figure 4 calls 3080/3092 a true-negative rate, although the matrix treats contaminated cases as positives; Table 41 also qualifies the main prose about repeated sampling.

Figure 4, p. 26; Section 3.5, p. 7; Table 41, p. 70 · Main paper

What to reuse—and what not to infer

Transferable lesson: Organize a large empirical paper around reproducible decisions, with compact takeaways linked to complete ablations, explicit held-out evaluation and safety checks.

Boundary: Do not generalize no-filtering wins to equal compute: that baseline uses twice as many examples. Preserve dataset-scale, model-family and safety limitations, and resolve main/appendix discrepancies before reusing numeric claims.

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: abstract, Sections 1–6 and acknowledgments. Official program title retained; proceedings PDF heading is Data Recipes for Reasoning Models. References excluded from analytical reading.

Appendix / supplement

Pages 20–72: contents and Appendices A–P, including all detailed generation/filtering prompts, ablations, licenses, safety and robustness analyses.

Visually inspected PDF pages

2, 9, 26

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Sun Mar 1 23:43:14 2026 EST · 72 pages