Writing study: AI Engram: In Search of Memory Traces in Artificial Intelligence

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Jea Kwon, Dong-Kyum Kim, Jiwon Kim, Yonghyun Kim, Woong Kook, Meeyoung Cha · 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: Theory & formal analysis (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

26 physical pages

Before references / appendix

≈ 9.21 page-equivalents

Main figures / tables

9 / 4

Attached appendix figures / tables

9 / 5

Attached appendix span

≈ 13.96 page-equivalents

Before references / appendix: ≈ 35.42%References / intervening material: ≈ 10.88%Attached appendix: ≈ 53.69%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Source-checked physical-page text restores wrapped top-level heading labels on pp. 14, 21, 25. Heading starts and measured positions are unchanged; this is label completeness, not a new section. Newly detected unnumbered impact/singular-acknowledgment statements are retained from the size-aware source extraction; authored heading corrections are unchanged.

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	≈ 0.9 pp.	3.46%
2. Related Work	2	≈ 0.47 pp.	1.81%
3. Defining Learning and Memory	2	≈ 0.64 pp.	2.46%
4. AI Engrams in Deep Neural Networks	3	≈ 2.04 pp.	7.85%
5. Experimental Validation	5	≈ 1.53 pp.	5.88%
6. Geometric Analysis of Engram	6	≈ 1.88 pp.	7.23%
7. In Search of Memory Trace in LLM	8	≈ 0.99 pp.	3.81%
8. Conclusion	9	≈ 0.28 pp.	1.08%
Impact Statement	10	≈ 0.11 pp.	0.42%
Acknowledgment	10	≈ 0.06 pp.	0.23%
References	10	≈ 2.83 pp.	10.88%
A. Derivation of Closed-form Solution	13	≈ 1.21 pp.	4.65%
B. Experimental Details for Causal Validation of Engram Method	14	≈ 2.37 pp.	9.12%
C. Surgical Validation of Engram Method.	16	≈ 4.42 pp.	17%
D. Operator-Theoretic Formalism of AI Engrams	21	≈ 0.5 pp.	1.92%
E. Experimental Setup for Large-Language Model Unlearning	21	≈ 3.1 pp.	11.92%
F. Compositional Memory States Hypothesis	24	≈ 0.72 pp.	2.77%
G. Structural Role of the Tabula Rasa Instantiation	25	≈ 1.18 pp.	4.54%
H. Limitations	26	≈ 0.46 pp.	1.77%

All detected figure and table locations

Main text

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

Attached appendix

• Table 5 · p. 15
• Figure 10 · p. 16
• Figure 11 · p. 16
• Figure 12 · p. 17
• Figure 13 · p. 18
• Figure 14 · p. 19
• Figure 15 · p. 20
• Figure 16 · p. 21
• Table 6 · p. 22
• Table 7 · p. 23
• Table 8 · p. 24
• Figure 17 · p. 24
• Figure 18 · p. 25
• Table 9 · p. 26

Argument and organization

1. Problem and contribution

   Makes four biological engram criteria operational through target/reference regions and observed/intervened model states.

2. Argument sequence

   Constraint formulation leads to the estimator and compositional editing, followed by validation, Fisher-geometry interpretation and a distinct LLM scaling test.

3. Evidence and verification

   The Fisher interpretation explicitly depends on isotropic output curvature, while the operational estimator does not; uniform edit strength is acknowledged to limit broader LLM performance.

Problem, claim and contribution

Makes four biological engram criteria operational through target/reference regions and observed/intervened model states.

Sections 3-4 and Table 1, pp. 2-3 · Main paper

Argument and section order

Constraint formulation leads to the estimator and compositional editing, followed by validation, Fisher-geometry interpretation and a distinct LLM scaling test.

Sections 4-7, pp. 3-9 · Main paper

Figures, captions and visual explanation

The state table separates natural learning from intervention; the LLM table keeps five metrics with direction arrows and shows utility/privacy tradeoffs alongside low exact memorization.

Table 1, p. 3; Table 3, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

The Fisher interpretation explicitly depends on isotropic output curvature, while the operational estimator does not; uniform edit strength is acknowledged to limit broader LLM performance.

Remarks 6.2-6.4 and Section 7, p. 8 · Main paper

Appendix and reproducibility

The reference-state appendix contrasts zero-state and fine-tuning-delta formulations, tests naive substitution and explicitly distinguishes functional decomposition from reconstruction of learning history.

Appendices G-H, pp. 25-26 · Main paper

Clarity and writing conventions

Memorable biological terminology is followed by mathematical criteria, but universal or surgical language is substantially narrowed by the final interpretive limitations.

Introduction, p. 2; Sections 5 and 7, pp. 5 and 8; Appendix H, p. 26 · Main paper

What to reuse—and what not to infer

Transferable lesson: Translate an analogy into testable constraints, then state which interpretation survives the practical approximation.

Boundary: The extracted component is retrospective and concept-dataset-dependent; it is not a recovered historical learning trace or a general guarantee of privacy-preserving unlearning.

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 1-8, pp. 1-9: biological criteria translated to constraints, retrospective estimator, unlearning tests, geometry and LLM limits.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-H, pp. 13-26: derivations, causal-validation setup, architecture-specific results, LLM metrics/costs, compositional hypothesis, reference-state choice and limitations.

Visually inspected PDF pages

3, 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. · 26 pages