Writing study: LLM DNA: Tracing Model Evolution via Functional Representations

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhaomin Wu, Haodong Zhao, Ziyang Wang, Jizhou Guo, Qian Wang, Bingsheng He · 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: Empirical analysis, diagnosis & auditing (primary) · 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

28 physical pages

Before references / appendix

≈ 10.49 page-equivalents

Main figures / tables

6 / 4

Attached appendix figures / tables

4 / 5

Attached appendix span

≈ 14.87 page-equivalents

Before references / appendix: ≈ 37.46%References / intervening material: ≈ 9.43%Attached appendix: ≈ 53.11%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	≈ 0.93 pp.	3.32%
2 RELATED WORK	2	≈ 0.71 pp.	2.54%
3 DEFINITION AND EXISTENCE OF LLM DNA	3	≈ 1.12 pp.	4%
4 EXTRACT LLM DNA	4	≈ 1.78 pp.	6.36%
5 EXPERIMENTS	6	≈ 4.69 pp.	16.75%
6 CONCLUSION	10	≈ 0.31 pp.	1.11%
REPRODUCIBILITY STATEMENT	11	≈ 0.39 pp.	1.39%
REFERENCES	11	≈ 2.64 pp.	9.43%
Appendix	14	≈ 0.5 pp.	1.79%
A PROOF	14	≈ 3.2 pp.	11.43%
B EXPERIMENTAL DETAILS	17	≈ 0.51 pp.	1.82%
C SUPPLEMENTARY EXPERIMENTS	18	≈ 4.25 pp.	15.18%
D BROADER IMPACT	22	≈ 0.77 pp.	2.75%
E LIMITATIONS	23	≈ 0.46 pp.	1.64%
F MODEL DETAILS	23	≈ 0.29 pp.	1.04%
G LARGE LANGUAGE MODEL USAGE	24	≈ 4.89 pp.	17.46%

All detected figure and table locations

Main text

• Figure 1 · p. 4
• Figure 2 · p. 7
• Figure 3 · p. 7
• Table 1 · p. 7
• Table 2 · p. 7
• Figure 4 · p. 8
• Table 3 · p. 9
• Table 4 · p. 9
• Figure 5 · p. 10
• Figure 6 · p. 10

Attached appendix

• Figure 7 · p. 19
• Table 5 · p. 20
• Figure 8 · p. 20
• Figure 9 · p. 20
• Table 1 · p. 20
• Figure 10 · p. 21
• Table 6 · p. 21
• Table 7 · p. 22
• Table 8 · p. 25

Argument and organization

1. DNA analogy and formal definition

   Translate inheritance and functional similarity into a bi-Lipschitz embedding proposal.

2. Practical extraction pipeline

   Separate semantic response encoding, sampled-input approximation and random projection.

3. Applications and stress tests

   Evaluate relation detection/routing before dataset, template, dimensionality and family analyses.

Problem, claim and contribution

The biological analogy is immediately tied to two stated properties, then formalized. This gives a memorable organizing device, but the theoretical inheritance statement is conditional on functional closeness rather than proof of historical descent.

Sections 1 and 3, pp. 1–4; Appendix A.1, pp. 14–15. · Main paper

Argument and section order

Section 4 explicitly names the two gaps between the abstract function-space definition and a usable algorithm: semantic response representation and finite prompt sampling. Its three subsections then resolve those implementation choices in order.

Section 4, pp. 4–5. · Main paper

Figures, captions and visual explanation

Figure 1 combines recognizable model/input icons with vector shapes and dimensions, showing response embedding, concatenation and projection in one left-to-right path. The caption is brief because labels already carry most of the mechanism.

Figure 1, p. 4. · Main paper

Tables, comparisons and evidence

Table 1 reports five-seed variability and multiple embedding encoders, while Table 2 separates downstream routing. The surrounding section identifies the 83 documented positive and 83 randomly sampled negative pairs rather than conflating 305 models with 305 independent labels.

Sections 5.2–5.3, pp. 6–8; Tables 1–2, p. 7. · Main paper

Appendix and reproducibility

A navigable contents page partitions proofs, settings, supplementary experiments, impacts and limitations. The full model inventory includes unknown licenses/architectures; the text also records quantization, decoding and chat-template choices needed to reproduce fingerprints.

Appendix contents, p. 14; B, pp. 17–18; E–G and Table 8, pp. 23–28. · Main paper

What to reuse—and what not to infer

Transferable lesson: Use an analogy to organize a definition, not replace it; explicitly identify each approximation between the formal object and the implemented measurement.

Boundary: Functional similarity and a midpoint-rooted distance tree do not prove training ancestry, license violations or a direction of evolution. The finite-set JL result does not itself certify that semantic embeddings preserve full logit-space behavior, and the sampled pair split is not a model-disjoint provenance test. Random-input robustness or nonsignificant size correlations are not universal invariance guarantees.

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

Sections 1–6, pp. 1–10; reproducibility/ethics and references pp. 11–13.

Appendix / supplement

Appendices A–G, pp. 14–28, including all proofs, extraction settings, robustness/ablation analyses, limitations and the 305-model inventory.

Visually inspected PDF pages

4, 7

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 28 pages