Writing study: In-the-Flow Agentic System Optimization for Effective Planning and Tool Use

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zhuofeng Li, Haoxiang Zhang, Seungju Han, Sheng Liu, Jianwen Xie, Yu Zhang, Yejin Choi, James Y Zou, Pan Lu · 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) · Architecture & representation design.

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

47 physical pages

Before references / appendix

≈ 10.23 page-equivalents

Main figures / tables

9 / 4

Attached appendix figures / tables

3 / 0

Attached appendix span

≈ 31.89 page-equivalents

Before references / appendix: ≈ 21.77%References / intervening material: ≈ 10.38%Attached appendix: ≈ 67.85%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	2	≈ 1.5 pp.	3.19%
2 PRELIMINARY	3	≈ 0.87 pp.	1.85%
3 IN-THE-FLOW AGENTIC SYSTEM OPTIMIZATION	4	≈ 2.12 pp.	4.51%
4 EXPERIMENTS	6	≈ 3.7 pp.	7.87%
5 RELATED WORK	10	≈ 0.47 pp.	1%
6 CONCLUSION	10	≈ 0.46 pp.	0.98%
REFERENCES	11	≈ 4.88 pp.	10.38%
A TRAINING ALGORITHM OF AGENTFLOW	16	≈ 1 pp.	2.13%
B THEORETICAL ANALYSIS OF FLOW-GRPO	17	≈ 2 pp.	4.26%
C EXPERIMENTAL DETAILS	19	≈ 3 pp.	6.38%
D MORE DISCUSSION ABOUT EXPERIMENT RESULTS	22	≈ 2 pp.	4.26%
E INSTRUCTION TEMPLATES IN AGENTFLOW	24	≈ 11 pp.	23.4%
F CASE STUDIES	35	≈ 12.89 pp.	27.43%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 3
• Figure 3 · p. 4
• Figure 4 · p. 5
• Table 1 · p. 7
• Table 2 · p. 7
• Figure 5 · p. 8
• Table 3 · p. 8
• Figure 6 · p. 9
• Figure 7 · p. 9
• Table 4 · p. 9
• Figure 8 · p. 9
• Figure 9 · p. 10

Attached appendix

• Figure 10 · p. 22
• Figure 11 · p. 22
• Figure 12 · p. 23

Argument and organization

1. Preliminary and AgentFlow

   Contrast monolithic tool policies with modular systems, then identify the planner as the trainable component.

2. Flow-GRPO and experiments

   Broadcast final outcome reward across planner turns and test accuracy, ablations, scaling and behavior.

3. Algorithm, theory and operational appendix

   Provide training pseudocode, assumptions, tool schemas, memory format, judge rubric and traces.

Argument and section order

The paper separates backbone replacement, offline SFT, on-policy RL, turn budget and tool upgrades. This makes several competing explanations inspectable instead of attributing everything to architecture.

Sections 4.3–4.6; Appendix D. · Main paper

Figures, captions and visual explanation

Figure 2 pairs the full loop with a magnified single transition. Consistent module colors and trained/frozen symbols distinguish optimization scope from runtime data flow.

Figure 2, page 3. · Main paper

Tables, comparisons and evidence

Table 3 and the adjacent before/after trace connect aggregate planner gains to a concrete recovery behavior. The SFT failure remains specific to this distillation setup, not evidence that supervised training generally fails.

Table 3 and Figure 5, page 8. · Main paper

The physics success trace explicitly contains a wrong time-dilation calculation yet ends with the reference option. This is an instructive mismatch between answer-level reward and process correctness.

Appendix F.6, pages 44–46. · Main paper

Appendix and reproducibility

The appendix supplies operational details absent from a model-size headline: training has three turns, evaluation ten; GPT-4o judges outcomes; retrieval uses an external embedding model and live web services.

Appendices C.1–C.2 and E.2–E.3. · Main paper

Clarity and writing conventions

Distinguish percentage-point gains from relative percentages and do not treat estimated proprietary parameter counts as established specifications. Sampled 100-question evaluations also need prominent scope labels.

Tables 1–2; Appendix C.4. · Main paper

What to reuse—and what not to infer

Transferable lesson: Make training ownership, runtime collaboration and evaluation access separately visible, and pair behavioral claims with trace-level evidence that survives checking.

Boundary: The appendix's surrogate decomposition does not by itself certify monotonic improvement for clipped, group-normalized, variable-length practical training. LLM-judged answers do not validate every intermediate action or computation.

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

Pages 1–10: Sections 1–6, all result and analysis panels; acknowledgments/references 11–14.

Appendix / supplement

Pages 15–47: contents, Appendices A–F, including derivations, training/evaluation, all prompts and seven complete case studies.

Visually inspected PDF pages

3, 8

• ICLR 2026 final proceedings · 47 pages