Writing study: MedAgentGym: A Scalable Agentic Training Environment for Code-Centric Reasoning in Biomedical Data Science

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Ran Xu, Yuchen Zhuang, Yishan Zhong, Yue Yu, Zifeng Wang, Xiangru Tang, Hang Wu, May Dongmei Wang, Peifeng Ruan, Donghan Yang, Tao Wang, Guanghua Xiao, Xin Liu, Carl Yang, Yang Xie, Wenqi Shi · 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: Datasets, benchmarks & measurement resources (primary) · Learning & optimization algorithms.

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

43 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

7 / 5

Attached appendix figures / tables

10 / 8

Attached appendix span

≈ 24.89 page-equivalents

Before references / appendix: ≈ 23.51%References / intervening material: ≈ 18.6%Attached appendix: ≈ 57.88%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	2	≈ 0.7 pp.	1.63%
2 RELATED WORKS	2	≈ 0.83 pp.	1.93%
3 MEDAGENTGYM: A SCALABLE AND INTERACTIVE LLM AGENT	3	≈ 2.05 pp.	4.77%
4 EVALUATING LLMS AS MEDICAL CODING AGENTS WITH MEDAGENTGYM	5	≈ 1.82 pp.	4.23%
5 TRAINING LLM AGENTS FOR CODE-CENTRIC BIOMEDICAL REASONING	7	≈ 2.87 pp.	6.67%
6 CONCLUSION	10	≈ 0.18 pp.	0.42%
ETHICS STATEMENT	10	≈ 0.19 pp.	0.44%
ACKNOWLEDGMENTS	10	≈ 0.36 pp.	0.84%
REFERENCES	11	≈ 8 pp.	18.6%
A LIMITATIONS AND BROADER IMPACTS	19	≈ 1.68 pp.	3.91%
B ADDITIONAL RELATED WORKS	20	≈ 1.4 pp.	3.26%
C TASK AND DATA DETAILS	22	≈ 3.27 pp.	7.6%
D BASELINE DETAILS	25	≈ 0.73 pp.	1.7%
E IMPLEMENTATION DETAILS	26	≈ 0.57 pp.	1.33%
F ADDITIONAL EXPERIMENTAL RESULTS	26	≈ 6.35 pp.	14.77%
G PROMPT DETAILS	33	≈ 10.89 pp.	25.33%

All detected figure and table locations

Main text

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

Attached appendix

• Table 6 · p. 20
• Figure 8 · p. 21
• Figure 9 · p. 22
• Figure 10 · p. 22
• Table 7 · p. 25
• Table 8 · p. 27
• Figure 11 · p. 27
• Figure 12 · p. 27
• Figure 13 · p. 27
• Table 9 · p. 28
• Table 10 · p. 28
• Figure 14 · p. 29
• Figure 15 · p. 29
• Figure 16 · p. 30
• Table 11 · p. 30
• Figure 17 · p. 31
• Table 12 · p. 31
• Table 13 · p. 32

Argument and organization

1. Two infrastructure gaps and benchmark construction

   Explains why both executable evaluation and an interactive training environment are needed.

2. Benchmarking and agent training

   Separates zero-shot diagnosis from SFT, offline/online RL and scaling experiments.

3. Generalization, failures and documentation

   Adds external suites, negative reward shaping, licensing and complete prompt templates.

Problem, claim and contribution

The introduction distinguishes domain knowledge from executable biomedical coding, making the infrastructure contribution more concrete than a generic medical-agent claim.

Section 1, p. 2 · Main paper

Argument and section order

Construction, environment, evaluation and training occupy separate sections, so readers can identify which artifact each experiment validates.

Sections 3–5 · Main paper

Figures, captions and visual explanation

Figure 2 uses color-coded data, agent, interface and sandbox regions, with a lower interaction strip translating the architecture into an episode.

Figure 2, p. 5 · Main paper

Tables, comparisons and evidence

Table 4 groups post-training variants under fixed backbones; external Table 5 and the unsuccessful repetition-penalty ablation prevent in-distribution improvement from standing alone.

Tables 4–5 and 13, pp. 8–9 and 32 · Main paper

Appendix and reproducibility

Per-source transformations, split rules, access licenses, task prompts and the human evaluation are documented individually instead of hidden behind one aggregate dataset count.

Appendices A.3, C, F.8 and G · Main paper

Clarity and writing conventions

Claims such as privacy-preserving or containerization guaranteeing compliance need narrower language: environment isolation, data permissions and external debugging-service routing are different controls.

Section 3.3; Appendices A.3 and E · Main paper

What to reuse—and what not to infer

Transferable lesson: For an environment paper, align each claimed capability with its own evidence: execution validity, training utility, generalization, cost and failure analysis.

Boundary: The prompts expose test-label directory placeholders, requiring an explicit leakage/access audit before reuse. Tables 3 and 4 report different averages for identical 7B backbone cells, and Table 8 turn counts exceed the stated 15-turn cap. Medical-looking case code is not independently validated clinical guidance.

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 and ethics

Appendix / supplement

Pages 19–43, Appendices A–G including all dataset transformations, training details, case studies and prompts

Visually inspected PDF pages

5, 8

• ICLR 2026 final proceedings · 43 pages