Writing study: Rational Transductors

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Mehryar Mohri · 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: Architecture & representation (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

36 physical pages

Before references / appendix

≈ 9.15 page-equivalents

Main figures / tables

6 / 1

Attached appendix figures / tables

3 / 1

Attached appendix span

≈ 22.91 page-equivalents

Before references / appendix: ≈ 25.42%References / intervening material: ≈ 10.94%Attached appendix: ≈ 63.64%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	≈ 1.51 pp.	4.19%
2. The Rational Features Framework	2	≈ 0.97 pp.	2.69%
3. Parameterization	3	≈ 0.99 pp.	2.75%
4. Expressivity and Complexity	4	≈ 1.42 pp.	3.94%
5. Theoretical Analysis of Learning	6	≈ 1.61 pp.	4.47%
6. Empirical Validation	7	≈ 1.58 pp.	4.39%
7. Conclusion	9	≈ 0.63 pp.	1.75%
Acknowledgments	10	≈ 0.04 pp.	0.11%
Impact Statement	10	≈ 0.07 pp.	0.19%
References	10	≈ 3.94 pp.	10.94%
A. Theoretical Background: Weighted Automata and Rational Power Series	14	≈ 2 pp.	5.56%
B. Extended Architectural Analysis	16	≈ 2 pp.	5.56%
C. Expressivity and Complexity	18	≈ 10 pp.	27.78%
D. Theoretical Analysis of Learning	28	≈ 7 pp.	19.44%
E. Additional Experiments and Details	35	≈ 1.66 pp.	4.61%
F. Limitations	36	≈ 0.25 pp.	0.69%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 7 · p. 17
• Figure 8 · p. 19
• Figure 9 · p. 25
• Table 2 · p. 35

Argument and organization

1. 1 Introduction; 2 Rational Features

   Introduces a matrix recurrence and layer-wise state injection.

2. 3 Parameterization; 4 Expressivity; 5 Learning

   Separates transition families, representational constructions and stability/generalization claims.

3. 6 Empirical Validation; 7 Conclusion

   Tests synthetic counting and arithmetic tasks and isolated mixing-layer latency before discussing scope.

Problem, claim and contribution

The central motivation is an inductive bias for state tracking; exact representability, learned finite-length behavior and full-language-model performance are different claims.

Introduction, p.1; Discussion and Scope, p.9 · Main paper

Argument and section order

The paper distinguishes conservation and contraction regimes, but contractive stability results should not be carried unchanged into the unit-gain counting regime.

Section 3.1, p.4; learning analysis, p.7 · Main paper

Figures, captions and visual explanation

The sidecar diagram makes independent projections into every block visible; the final page places isolated latency and two distinct arithmetic metrics in separate panels.

Figure 1, p.3; Figures 5-6, p.9 · Main paper

Tables, comparisons and evidence

The extrapolation baseline uses absolute-position entries left at random initialization beyond training length, while the latency test compares isolated sequence mixers rather than the complete hybrid model.

Sections 6.2-6.3, p.8 · Main paper

Appendix and reproducibility

The supplement reports five-seed results, explains normalized Float64 arithmetic targets and records the state-dimension versus scan-cost limitation.

Appendices E.4-E.5 and F, p.36 · Main paper

Clarity and writing conventions

Strong claims of exact learned algorithms need qualification: the main text reports finite-precision drift and expressly omits large-scale NLP pretraining.

Footnote 1, p.8; Discussion and Scope, p.9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Keep architecture diagrams precise about retained components, and distinguish expressive constructions from learned solutions and component-level speed tests.

Boundary: The stated Transformer separations require careful model and precision assumptions; AC0 and TC0 are discussed differently within the paper, and this writing scan does not certify the theoretical claims or infer general Transformer failure from the chosen baseline.

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-oriented scan of the main argument through physical p.9, with focused physical-page text inspection on 1, 3, 4, 5, 7, 8, 9. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 13, 20, 36. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

3, 9

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 proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 36 pages