Writing study: TG-RAG: A Retrieval-Augmented Framework for Reasoning Guidance in Specialized Domains

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Liang Su, Mingyang Zhang, Yun Xiong, Tengfei Liu, Siwei Zhang, Xi Chen, Li Sun · 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: Domain-grounded methods & physical systems (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

17 physical pages

Before references / appendix

≈ 9.04 page-equivalents

Main figures / tables

4 / 4

Attached appendix figures / tables

4 / 2

Attached appendix span

≈ 4.91 page-equivalents

Before references / appendix: ≈ 53.18%References / intervening material: ≈ 17.94%Attached appendix: ≈ 28.88%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	≈ 1.29 pp.	7.59%
2. Related Work	2	≈ 0.72 pp.	4.24%
3. Methodology	3	≈ 2.69 pp.	15.82%
4. Experiment	6	≈ 3.6 pp.	21.18%
5. Conclusion	9	≈ 0.17 pp.	1%
Acknowledgements	9	≈ 0.06 pp.	0.35%
Impact Statement	9	≈ 0.17 pp.	1%
References	10	≈ 3.04 pp.	17.88%
A. Empirical Analysis of Open Branch Activation	13	≈ 0.17 pp.	1%
B. EPG Construction and Real-World Example	13	≈ 0.96 pp.	5.65%
C. Efficiency Analysis	14	≈ 0.62 pp.	3.65%
D. Detailed Prompt Description	14	≈ 0.58 pp.	3.41%
E. Baseline Implementation and Knowledge Source Details	15	≈ 2.58 pp.	15.18%

All detected figure and table locations

Main text

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

Attached appendix

• Table 5 · p. 15
• Table 6 · p. 15
• Figure 5 · p. 16
• Figure 6 · p. 17
• Figure 7 · p. 17
• Figure 8 · p. 17

Argument and organization

1. Problem and contribution

   Distinguishes retrieved facts from procedural directives, presenting static context adherence as the gap addressed by stepwise graph routing.

2. Argument sequence

   The graph's state/directive/context representation precedes interruption and dual-path injection; ablations separately test branching, information channels and placement of guidance.

3. Evidence and verification

   The control variants hold the expert-procedure source while changing fixed versus conditional routes or prompt versus process placement; expert graph review is a disclosed construction cost.

Problem, claim and contribution

Distinguishes retrieved facts from procedural directives, presenting static context adherence as the gap addressed by stepwise graph routing.

Introduction, p. 2 · Main paper

Argument and section order

The graph's state/directive/context representation precedes interruption and dual-path injection; ablations separately test branching, information channels and placement of guidance.

Sections 3-4.3, pp. 3-8 · Main paper

Figures, captions and visual explanation

A two-lane static/dynamic diagram exposes interruption, retrieval and resumed generation, with injected directives highlighted; neighboring tables isolate graph structure from integration strategy.

Figure 2, p. 3; Tables 2-3, p. 8; both PNGs inspected · Main paper

Tables, comparisons and evidence

The control variants hold the expert-procedure source while changing fixed versus conditional routes or prompt versus process placement; expert graph review is a disclosed construction cost.

Section 4.3, p. 8; Appendix B.1, p. 13 · Main paper

Appendix and reproducibility

The supplement supplies graph representation, continuation/router/fallback prompts and approximately 30-minute task-class expert review, exposing dependencies hidden by the training-free label.

Appendix B, pp. 13-14; Figures 6-8, p. 17 · Main paper

Clarity and writing conventions

Repeated facts-versus-steps wording makes the design legible, but terms such as enforceable or rigidly constrained are stronger than probabilistic text conditioning guarantees.

Sections 3.3.2-3.3.3, p. 5 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate knowledge, procedure and control flow, then design ablations that change one integration choice at a time.

Boundary: Low fallback frequency is not proof of out-of-distribution safety, and benchmark medical/financial accuracy is not clinical or compliance validation.

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: procedural versus factual guidance, graph representation, interruption/routing/injection, task comparisons and component controls.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendices A-E and prompt figures, pp. 13-17: fallback frequency, graph construction and expert review, cost, knowledge-source protocol and exact prompts.

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. · 17 pages