Writing study: Shoot First, Ask Questions Later? Building Rational Agents that Explore and Act Like People

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Gabriel Grand, Valerio Pepe, Joshua B Tenenbaum, Jacob Andreas · 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: Learning & optimization algorithms (primary) · Empirical analysis, diagnosis & auditing · Datasets, benchmarks & measurement resources.

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

46 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

5 / 1

Attached appendix figures / tables

15 / 8

Attached appendix span

≈ 28.8 page-equivalents

Before references / appendix: ≈ 21.98%References / intervening material: ≈ 15.41%Attached appendix: ≈ 62.61%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	≈ 1.48 pp.	3.22%
2 THE BATTLESHIP GAME	3	≈ 0.37 pp.	0.8%
3 FORMAL FRAMEWORK: BAYESIAN EXPERIMENTAL DESIGN	3	≈ 1.71 pp.	3.72%
4 EXPERIMENTS	5	≈ 3.54 pp.	7.7%
5 GENERALIZED INFORMATION-SEEKING GAMES	8	≈ 0.39 pp.	0.85%
6 RELATED WORK, DISCUSSION, AND CONCLUSION	9	≈ 1 pp.	2.17%
AUTHOR CONTRIBUTIONS	10	≈ 0.23 pp.	0.5%
ACKNOWLEDGMENTS	10	≈ 0.2 pp.	0.43%
ETHICS STATEMENT	10	≈ 0.15 pp.	0.33%
REPRODUCIBILITY STATEMENT	10	≈ 0.41 pp.	0.89%
REFERENCES	11	≈ 7.1 pp.	15.43%
Appendix	18	≈ 0.9 pp.	1.96%
A BATTLESHIP	19	≈ 9 pp.	19.57%
B SEQUENTIAL MONTE CARLO APPROXIMATION	28	≈ 1 pp.	2.17%
C PROMPT LIBRARY	29	≈ 4 pp.	8.7%
D EVALUATIONS	33	≈ 13.89 pp.	30.2%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 20
• Table 2 · p. 22
• Figure 7 · p. 23
• Figure 8 · p. 24
• Figure 9 · p. 24
• Figure 10 · p. 25
• Figure 11 · p. 25
• Figure 12 · p. 25
• Figure 13 · p. 26
• Figure 14 · p. 26
• Figure 15 · p. 26
• Figure 16 · p. 27
• Table 3 · p. 33
• Figure 17 · p. 34
• Table 4 · p. 38
• Figure 18 · p. 38
• Figure 19 · p. 39
• Table 5 · p. 39
• Table 6 · p. 40
• Table 7 · p. 41
• Table 8 · p. 43
• Figure 20 · p. 43
• Table 9 · p. 44

Argument and organization

1. Controlled game and Bayesian formalization

   Turn broad information-seeking claims into distinct asking, answering and action decisions.

2. Human study then agent decomposition

   Establish annotated behavior before testing the Spotter and Captain roles.

3. Replication and appendix

   Test a second game and retain protocols, prompts and full results.

Problem, claim and contribution

The game creates a measurable exploration/action tradeoff with a binary answer bottleneck; this is a narrower and testable claim than general rational agency.

Sections 1–2, pp. 1–3 · Main paper

Figures, captions and visual explanation

Figure 1 combines visible/hidden board views with Captain/Spotter roles and computational thought bubbles. It explains both the human task and the inference machinery in one scene.

Figure 1, p. 2 · Main paper

Tables, comparisons and evidence

Figure 4 separates targeting F1, question information gain and question timing. The adjoining result explicitly notes that better questions alone need not improve gameplay.

Figure 4 and Section 4.3.1, pp. 7–8 · Main paper

Appendix and reproducibility

A contents page navigates human recruitment/incentives, annotation definitions, selected and random examples, SMC equations, exact prompts, full metrics, token costs and a second game.

Appendices A–E, pp. 18–46 · Main paper

Clarity and writing conventions

Short intuition paragraphs immediately follow the Bayesian update and information-gain equations, preserving rigor while explaining what each quantity means.

Section 3, pp. 3–5 · Main paper

What to reuse—and what not to infer

Transferable lesson: Decompose an agent claim into the ability to obtain information and the ability to use it. Pair formal equations with plain-language intuition, and let results that break the intuitive story remain visible.

Boundary: Check text/caption consistency before copying a design: the main Spotter analysis names Mann–Whitney U, whereas Figure 17 names a paired t-test; Table 4 prints epsilon=1.0 despite the stated 0.1 noise model. Game-specific success and Captain token costs are not general real-world autonomy or complete system cost.

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

PDF pp. 1–10, Sections 1–6 and ethics/reproducibility/AI-assistance statements; reference list excluded from analytical reading

Appendix / supplement

Appendices A–E, pp. 18–46: human study, annotations/examples, SMC, full prompts/programs, evaluations/costs and Guess Who replication

Visually inspected PDF pages

2, 8

• Official ICLR 2026 proceedings conference PDF; PDF creation metadata: Thu Feb 26 16:31:09 2026 EST · 46 pages