Writing study: ExComm: Exploration-Stage Communication for Error-Resilient Agentic Test-Time Scaling

NeurIPS overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← NeurIPS 2026 oral writing corpus

Woomin Song, Beomjun Kim, Daewon Choi, Sai Muralidhar Jayanthi, Saket Dingliwal, Jinwoo Shin, Aram Galstyan · Official NeurIPS 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

17 physical pages

Before references / appendix

≈ 9.09 page-equivalents

Main figures / tables

2 / 6

Attached appendix figures / tables

2 / 1

Attached appendix span

≈ 5.91 page-equivalents

Before references / appendix: ≈ 53.47%References / intervening material: ≈ 11.76%Attached appendix: ≈ 34.76%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Top-level heading and reference/appendix boundaries checked against source PDF text. Bibliography titles/author initials, numbered contribution/property bullets, diagram steps, equations and TOC entries are not substantive headings. Section spans are page-coordinate estimates, not ink area or prose-only length. This is an author manuscript mirror, not a verified accepted-version template. Source layout uses physical single-column y/page-height, including equation-heavy pages.

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	≈ 1.64 pp.	9.65%
2 Related Works	3	≈ 0.31 pp.	1.82%
3 ExComm	3	≈ 2.97 pp.	17.47%
4 Experiments	6	≈ 3.16 pp.	18.59%
5 Conclusion	9	≈ 0.32 pp.	1.88%
References	10	≈ 2 pp.	11.76%
A Further Discussions	12	≈ 0.45 pp.	2.65%
B Implementation Details on Critic Model	12	≈ 1.71 pp.	10.06%
C Error Type Analysis	14	≈ 0.45 pp.	2.65%
D Latency Analysis	14	≈ 1.39 pp.	8.18%
E Example Illustrations	16	≈ 1.18 pp.	6.94%
F Case Studies	17	≈ 0.73 pp.	4.29%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 3 · p. 13
• Figure 4 · p. 15
• Table 7 · p. 15

Argument and organization

1. Problem and contribution

   Targets errors propagating during exploration, distinguishing early correction from post-hoc trajectory selection or refinement.

2. Argument sequence

   Formalizes agent beliefs and plans before consistency checks, diversification, benchmark results, recovery analysis and cost.

3. Evidence and verification

   The ablation retains a non-winning intermediate variant; the cost plot reports API spending separately from the appendix's wall-clock latency.

Problem, claim and contribution

Targets errors propagating during exploration, distinguishing early correction from post-hoc trajectory selection or refinement.

Introduction · Main paper

Argument and section order

Formalizes agent beliefs and plans before consistency checks, diversification, benchmark results, recovery analysis and cost.

Sections 3–4 · Main paper

Figures, captions and visual explanation

The architecture gives belief consistency and plan diversification separate colored paths, then shows their different updates inside a repeated exploration loop.

Figure 1, p. 2 · Main paper

Tables, comparisons and evidence

The ablation retains a non-winning intermediate variant; the cost plot reports API spending separately from the appendix's wall-clock latency.

Table 6; Figure 2, p. 9; Table 7 · Main paper

Appendix and reproducibility

Literal critic prompts, error-type definitions, resolution objects and case studies make the communication mechanism inspectable.

Appendices B–F · Main paper

Clarity and writing conventions

The paper names limited GAIA subsets and the evaluation tools instead of treating benchmark labels as a sufficient protocol description.

Table 2 caption; Section 4 · Main paper

What to reuse—and what not to infer

Transferable lesson: For agent systems, diagram mutable state and intervention timing, then test failure recovery and resource costs separately.

Boundary: The critic-based recovery measure is not an independently verified ground-truth trace annotation. Agent-count matching does not alone match total tokens, latency or spend.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-02

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-oriented distributed manuscript scan of Sections 1–5, pp. 1–9; references pp. 10–11: abstract/framing, section sequence, substantive text windows, captions, evidence organization and conclusion; not exhaustive technical verification.

Appendix / supplement

Writing-oriented scan of Appendices A–F, pp. 12–17; critics, error categories, latency and worked trajectories; code and experiments were not executed.

Visually inspected PDF pages

2, 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.

• Public author arXiv manuscript 2605.22102v1; not verified identical to the accepted or camera-ready OpenReview revision · 17 pages