Writing study: FLIP2: Expanding Protein Fitness Landscape Benchmarks for Real-World Machine Learning Applications

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Kieran Didi, Sarah Alamdari, Alex Xijie Lu, Bruce James Wittmann, Kadina E Johnston, Ava P Amini, Ali Madani, Maya Czeneszew, Christian Dallago, Kevin K Yang · Official ICML 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 (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

54 physical pages

Before references / appendix

≈ 8.57 page-equivalents

Main figures / tables

4 / 1

Attached appendix figures / tables

32 / 18

Attached appendix span

≈ 41.96 page-equivalents

Before references / appendix: ≈ 15.87%References / intervening material: ≈ 6.43%Attached appendix: ≈ 77.7%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	≈ 0.78 pp.	1.44%
2. Related Work	2	≈ 1.43 pp.	2.65%
3. The FLIP2 Datasets and Splits	3	≈ 2.49 pp.	4.61%
4. Evaluations	6	≈ 2.38 pp.	4.41%
5. Discussion	8	≈ 0.51 pp.	0.94%
Impact Statement	9	≈ 0.35 pp.	0.65%
References	9	≈ 3.47 pp.	6.43%
A. Supplemental Tables and Figures	13	≈ 41.96 pp.	77.7%

All detected figure and table locations

Main text

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

Attached appendix

• Table A1 · p. 13
• Table A2 · p. 14
• Table A3 · p. 14
• Table A4 · p. 15
• Table A5 · p. 15
• Table A6 · p. 16
• Table A7 · p. 16
• Table A8 · p. 17
• Table A9 · p. 17
• Table A10 · p. 18
• Table A11 · p. 18
• Table A12 · p. 19
• Table A13 · p. 19
• Table A14 · p. 20
• Table A15 · p. 20
• Table A16 · p. 21
• Table A17 · p. 21
• Table A18 · p. 22
• Figure A1 · p. 23
• Figure A2 · p. 24
• Figure A3 · p. 25
• Figure A4 · p. 26
• Figure A5 · p. 27
• Figure A6 · p. 28
• Figure A7 · p. 29
• Figure A8 · p. 30
• Figure A9 · p. 31
• Figure A10 · p. 32
• Figure A11 · p. 33
• Figure A12 · p. 34
• Figure A13 · p. 35
• Figure A14 · p. 36
• Figure A15 · p. 37
• Figure A16 · p. 38
• Figure A17 · p. 39
• Figure A18 · p. 40
• Figure A19 · p. 41
• Figure A20 · p. 42
• Figure A21 · p. 43
• Figure A22 · p. 44
• Figure A23 · p. 45
• Figure A24 · p. 46
• Figure A25 · p. 47
• Figure A26 · p. 48
• Figure A27 · p. 49
• Figure A28 · p. 50
• Figure A29 · p. 51
• Figure A30 · p. 52
• Figure A31 · p. 53
• Figure A32 · p. 54

Argument and organization

1. Problem and contribution

   Motivates held-out wild types, positions and mutations as distinct engineering generalization demands rather than treating a random train/test split as sufficient.

2. Argument sequence

   Dataset identity, split type and baseline family are introduced as separate axes, then evaluation progresses from zero-shot scores to linear models to fine-tuned models.

3. Evidence and verification

   Fine-tuning uses five random seeds and is contrasted with pretrained and random-weight models; the discussion retains regimes in which simpler baselines win.

Problem, claim and contribution

Motivates held-out wild types, positions and mutations as distinct engineering generalization demands rather than treating a random train/test split as sufficient.

Introduction and Sections 3.1-3.2, pp. 1-5 · Main paper

Argument and section order

Dataset identity, split type and baseline family are introduced as separate axes, then evaluation progresses from zero-shot scores to linear models to fine-tuned models.

Figure 1, p. 2; Section 4, pp. 6-8 · Main paper

Figures, captions and visual explanation

The overview uses three labeled horizontal bands for datasets, splits and baselines. Paired scatter plots use equality diagonals and consistent split-type colors to reveal both positive and negative transfer.

Figure 1, p. 2; Figure 4, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

Fine-tuning uses five random seeds and is contrasted with pretrained and random-weight models; the discussion retains regimes in which simpler baselines win.

Section 4.3 and discussion, pp. 8-9 · Main paper

Appendix and reproducibility

Per-split tables retain both Spearman and NDCG, including undefined correlation entries; extensive prediction panels distinguish model families and note seed averaging.

Tables A1-A3, pp. 13-14; Figures A8 and A32, pp. 30 and 54 · Main paper

Clarity and writing conventions

The findings emphasize conditional utility across split types rather than declaring one protein language model universally best.

Section 5, pp. 8-9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use parity plots and split-specific tables to show when an expensive intervention fails, not only its aggregate winning score.

Boundary: Different held-out regimes test different capabilities; do not pool them into a single biological-generalization claim or replace undefined correlations with zero.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-5, pp. 1-9: engineering-motivated benchmark expansion, split definitions, datasets, baseline families and comparative conclusions.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendix A, pp. 13-54: per-split metric tables and aligned prediction panels, including continued figure pages and seed-averaged fine-tuned predictions.

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.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 54 pages