Writing study: mCLM: A Modular Chemical Language Model that Generates Functional and Makeable Molecules

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Carl Edwards, Chi Han, Gawon Lee, Thao Nguyen, Sara Szymkuć, Chetan Prasad, Bowen Jin, Jiawei Han, Ying Diao, Ge Liu, Hao Peng, Bartosz Grzybowski, Martin Burke, Heng Ji · 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: Architecture & representation design (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

33 physical pages

Before references / appendix

≈ 11.11 page-equivalents

Main figures / tables

5 / 2

Attached appendix figures / tables

8 / 9

Attached appendix span

≈ 13.89 page-equivalents

Before references / appendix: ≈ 33.67%References / intervening material: ≈ 24.24%Attached appendix: ≈ 42.09%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	≈ 3.07 pp.	9.3%
2 THE MODULAR CHEMICAL LANGUAGE MODEL	4	≈ 2.49 pp.	7.55%
3 EXPERIMENTAL EVALUATION	7	≈ 2.33 pp.	7.06%
4 RELATED WORK	9	≈ 0.82 pp.	2.48%
5 CONCLUSIONS AND FUTURE WORK	10	≈ 0.26 pp.	0.79%
REPRODUCIBILITY STATEMENT	10	≈ 0.12 pp.	0.36%
ACKNOWLEDGMENTS	10	≈ 1.3 pp.	3.94%
REFERENCES	12	≈ 8 pp.	24.24%
A ADDITIONAL RESULTS	20	≈ 1.56 pp.	4.73%
B CRITICAL CHEMICAL REASONING ALGORITHM	21	≈ 0.89 pp.	2.7%
C PRETRAINING DATA MIXTURE	22	≈ 1.24 pp.	3.76%
D SYNTHETIC DATA GENERATION	23	≈ 2.64 pp.	8%
E TRAINING PROCEDURE	26	≈ 0.68 pp.	2.06%
F TOKENIZER DETAILS	27	≈ 4 pp.	12.12%
G SOURCE DATASETS AND DATABASES	31	≈ 1.33 pp.	4.03%
H ADDITIONAL BASELINES	32	≈ 1.05 pp.	3.18%
I ABLATION STUDIES	33	≈ 0.51 pp.	1.55%

All detected figure and table locations

Main text

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

Attached appendix

• Table 3 · p. 20
• Table 4 · p. 21
• Figure 6 · p. 21
• Figure 7 · p. 22
• Table 5 · p. 23
• Table 6 · p. 24
• Table 7 · p. 26
• Figure 8 · p. 28
• Figure 9 · p. 28
• Figure 10 · p. 29
• Figure 11 · p. 29
• Figure 12 · p. 30
• Figure 13 · p. 30
• Table 8 · p. 32
• Table 9 · p. 33
• Table 10 · p. 33
• Table 11 · p. 33

Argument and organization

1. Representation matched to physical synthesis

   Links the tokenization decision to a domain constraint.

2. Vocabulary, model, iterative reasoning

   Separates three method levels before evaluation.

3. Property editing and makeability

   Evaluates predicted function and synthesis feasibility separately, then illustrates trade-offs in case studies.

Problem, claim and contribution

The word-versus-character analogy makes block tokenization accessible, but the long introduction's societal and human-reasoning claims dilute the specific technical argument.

Sections 1–2, pp. 1–4 · Main paper

Argument and section order

The molecule-editing case study exposes property trade-offs across iterations, which communicates multi-objective optimization more effectively than a single final molecule.

Figure 4, p. 9 · Main paper

Figures, captions and visual explanation

The tokenizer diagram shows a coverage-check branch and fallback, so an important exception to synthesis-guaranteed tokenization is visible rather than buried in prose.

Figure 3, p. 5 · Main paper

Tables, comparisons and evidence

Table 2 defines validity, conditional synthesizability and their product separately, helping readers distinguish a valid string from a predicted usable synthesis route.

Table 2, p. 8 · Main paper

Appendix and reproducibility

The reproducibility statement points readers to oracle construction, data mixtures, training stages and tokenizer details; Tables 5–6 separate collected data from the much smaller trained subset.

Reproducibility statement, p. 10; Appendices C–G, pp. 22–32 · Main paper

Clarity and writing conventions

The main-text unified loss description needs qualification from Appendix E's separately averaged modality losses; Algorithm 1 also omits an explicit candidate-state update and finite iteration bound.

Section 2.2, p. 6; Algorithm 1, p. 21; Appendix E, p. 26 · Main paper

What to reuse—and what not to infer

Transferable lesson: Let the domain constraint determine both representation and evaluation, and diagram exceptions or fallback paths alongside the idealized method.

Boundary: Predicted ADMET improvements and computational retrosynthesis are not validated clinical improvements or experimental synthesis of the proposed molecules. Guarantees depend on restricted blocks and reaction assumptions; Table 2's numbers also differ from its following MoleculeSTM prose.

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

PDF pp. 1–11, Sections 1–5 and reproducibility statement

Appendix / supplement

PDF pp. 20–33, Appendices A–I

Visually inspected PDF pages

5, 8

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Sun Mar 1 12:51:54 2026 EST · 33 pages