The paper has a GitHub, the GitHub has a

README, the README has nothing:

Reproducibility Signals for Review Support

Federico Bolelli

⋆, B

, Davide Santoli

⋆

, Kevin Marchesini,

Luca Lumetti, and Costantino Grana

University of Modena and Reggio Emilia, Italy

{name.surname}@unimore.it

Abstract. Reproducibility policies promise “checkable” medical-imaging

science, yet many submissions still ship unveriﬁable artifacts. Our analy-

sis of 3 722 miccai papers shows code-linking rising from 51.8% (2021) to

72.5% (2025), but ∼13% of linked repositories are inaccessible or empty.

We present paper-snitch, a reviewer-facing decision-support tool that

turns these signals into an evidence-grounded report. paper-snitch pars-

es pdfs, resolves and sanity-checks repositories, and applies policy-aware

checklists aligned with miccai expectations, producing a review-time ver-

iﬁability score decomposed into interpretable sub-scores plus criterion-

linked excerpts and artifacts reviewers can inspect. It never executes

untrusted code or attempts gpu-heavy reproduction, focusing instead

on bounded, veriﬁable checks. We compare paper-snitch on 100 ran-

domly sampled miccai 2025 papers with human annotators using shared

evaluation criteria, indicating that automated, bounded checks can scale

reproducibility screening while keeping ﬁnal decisions with reviewers.

Keywords: Reproducibility Screening · Artifact Veriﬁcation · Peer-review

Support Tools · paper-snitch

1 Introduction

Let us start by lowering your blood pressure: this paper does not introduce

a new transformer variant, a clever loss, or a “lightweight” architecture that

beats the state of the art (with a standard deviation, naturally, omitted for

space constraints). Instead, we focus on a less glamorous but increasingly urgent

problem for the miccai community: how to keep our science checkable when the

volume and complexity of scientiﬁc production are all accelerating at once.

The current wave of (medical imaging) submissions is shaped by two trends:

rapidly growing pipeline complexity and the widespread use of llm-based code

and text generation. While these tools can streamline engineering and documen-

tation, they also amplify risks—overconﬁdent narratives around fragile exper-

iments, incomplete disclosure, and, in the worst case, low-quality manuscripts

⋆

Equal contribution. Authors are allowed to list their names ﬁrst on their CVs.

B Corresponding author: federico.bolelli@unimore.it

2 F. Bolelli, D. Santoli, et al.

(a) Code availability in miccai proceedings (b) Findings

Fig. 1: (a) Analysis of 3 722 papers showing total publications (blue), papers

with code links (green), and papers with veriﬁed accessible code (red). Code

sharing has increased from 51.8% (2021) to 72.5% (2025), though ∼13% of linked

repositories remain inaccessible or empty. Overall, 65.5% provide links, only

54.1% have actual code (see Sec. 3.2). (b) Evidence-highlighted paper view with

criterion-linked excerpts.

that strain peer review [6,19]. Reviewers are asked to judge technical novelty,

clinical relevance, evaluation rigor, and ethical considerations under strict time

constraints, while reproducibility requirements add another axis of responsibility.

Reproducibility policies exist precisely to make this feasible. miccai provides

explicit guidance [18] on what should be made available to support reproducible

research (e.g., code, trained models where appropriate, documentation of data

and evaluation, and clear statements describing what is and is not released).

Other computer vision and pattern-recognition venues increasingly complement

such guidance with badges or artifact evaluation tracks [1,10] that recognize

veriﬁable releases and incentivize authors to produce them. The intent is clear:

shift the norm from “trust me” to “here is what you need to check.”

Yet, in practice, the weakest link is not the policy but the workﬂow (Fig. 1a):

many papers still provide no code, or promise a release “after acceptance” that

never materializes in the camera-ready [2,8]. Others provide a repository link

that is empty, private, missing critical components, or contains little more than

a placeholder README [3,13,24]. Even when artifacts are provided, a reviewer

cannot realistically execute arbitrary code, resolve dependencies, acquire datasets,

or reproduce gpu-heavy training. Consequently, compliance becomes diﬃcult to

assess, and the boundary between “not veriﬁable within review constraints” and

“not suﬃciently supported by provided artifacts” becomes blurred.

Crucially, a substantial part of this problem is mechanical and therefore

amenable to automation: many review-time questions are objective veriﬁcations

rather than scientiﬁc judgments. For example, does a linked repository exist and

contain runnable code beyond a stub? Is there an explicit license? Are environ-

ment speciﬁcations provided (e.g., requirements.txt, environment.yml)? Are

Reproducibility Signals for Review Support 3

pretrained weights and evaluation scripts available, and are paper claims trace-

able to concrete commands? Are data access instructions, splits, and evaluation

protocols clearly speciﬁed, and do the artifacts match the venue’s reproducibil-

ity statement? This motivates our central question: can we leverage modern

AI tooling to perform these bounded checks and present the results as decision

support—not an automated gatekeeper—in a transparent, evidence-grounded re-

port that highlights what is present, what is missing or unveriﬁable, and what

still requires human interpretation?

In this work, we introduce paper-snitch, a tool that supports review-time

screening of reproducibility signals for medical imaging submissions by combin-

ing lightweight repository inspection, structured text analysis of the manuscript,

and policy-aware checklists aligned with miccai’s stated expectations [18]. The

tool outputs (a) a review-time veriﬁability score (artifact completeness and con-

sistency) decomposed into interpretable sub-scores, and (b) a set of grounded

ﬁndings that can be inspected by the reviewer (Fig. 1b). The workﬂow is de-

signed to be scalable and to integrate into a venue review pipeline as an auxiliary

report, similar in spirit to artifact evaluation initiatives, but tailored to the time

and safety constraints of double-blind peer review. It does not establish end-to-

end reproducibility; it only checks what can be veriﬁed from provided artifacts

under review-time constraints.

We start from miccai because its scope makes reproducibility both essential

and challenging. Nonetheless, the approach is modular and can be adapted to

other (computer vision) venues with diﬀerent policies or requirements [1,10].

If this already convinced the reader, the tool implementation is available on

GitHub,

while the live version is hosted at https://paper-snitch.unimore.it.

2 Related Work

Reproducible Research, Checklists, and Artifacts. Reproducible compu-

tational science increasingly treats the contribution as a research compendium—

paper, code, data (when shareable), and instructions to reproduce results [23].

Empirical studies report that many AI papers omit key details or artifacts

needed to reproduce ﬁndings [5]. In response, major venues have introduced

reproducibility checklists and artifact evaluation/badging to incentivize and sig-

nal veriﬁable releases [1,18,10,20,21]. However, veriﬁcation remains costly and

inconsistent, especially under review-time constraints.

Reproducibility in Medical Imaging. Biomedical image analysis faces ad-

ditional barriers: restricted patient data, heterogeneous acquisition, and prepro-

cessing choices that can dominate outcomes. Reporting standards and commu-

nity guidance emphasize documenting data handling and evaluation protocols to

improve interpretability and reproducibility [16]. Similar issues arise in miccai-

style methodological work, where seemingly minor implementation diﬀerences

can substantially change results while remaining underreported.

https://github.com/AImageLab-zip/paper-snitch

paper-snitch makes use of paid services; unauthenticated users can only visualize.

4 F. Bolelli, D. Santoli, et al.

Fig. 2: End-to-end workﬂow of paper-snitch. Yellow nodes denote llm calls;

conditional branches are highlighted in red (diamonds and labels).

Automation for Review-time Screening. Recent discussions on generative

AI in peer review highlight both opportunities for scalable consistency checks

and risks around privacy, bias, and reliability [6,19]. Many reproducibility criteria

are largely veriﬁable (artifact presence, environment specs, runnable evaluation,

claim-to-command traceability) rather than requiring full re-execution. Our work

targets this veriﬁable surface layer with policy-aware, evidence-grounded reports

designed to support—not replace—human judgment.

3 Our Solution

3.1 Design Goals

paper-snitch is designed as decision support for review-time reproducibility

screening (artifact veriﬁability), not an automated acceptance ﬁlter. This fram-

ing drives three non-negotiable constraints: (i) evidence-grounded outputs, every

claim made by the tool must be traceable to an inspectable artifact (pdf text

span, repository ﬁle, or link), (ii) review-time safety, the system never executes

untrusted code and never attempts to reproduce gpu-heavy training runs. In-

stead, it performs static inspection, and (iii) scalability and auditability, the

workﬂow must handle many papers concurrently, support partial completion,

expose per-node failures, and preserve detailed intermediate artifacts so that

chairs or a dedicated reproducibility track can audit how scores were produced.

3.2 System Overview

Reproducibility screenings are operationalized through the workﬂow in Fig. 2.

The output is a compact report with decomposed scores and per-criterion evi-

dence snippets that a reviewer can inspect. Nodes are detailed as follows.

Paper Parsing. Given a pdf, we extract a structured representation (tei/xml)

of the manuscript using grobid [15], a machine learning library focused on

Reproducibility Signals for Review Support 5

technical and scientiﬁc publications. The result is stored as section-wise text

(and, when possible, ﬁgs./tabs. content and captions), so downstream steps can

reference evidence at the level of concrete spans rather than free-form summaries.

Paper Classiﬁcation. In parallel, we classify each paper as introducing a new

dataset, proposing a new method, addressing both (dataset & method), or being

purely theoretical, i.e., not requiring any code. This categorization is performed

by a llm and leverages the full content of the paper. This label has two explicit

eﬀects in the workﬂow: it determines which branches are applicable (e.g., pure

theoretical papers can skip code analysis checks), and it selects the weighting

used to compute the ﬁnal score, which depends on the paper’s category. What

constitutes missing reproducibility information is contribution-dependent: the

system should avoid penalizing inapplicable items.

Section Embedding. The manuscript is indexed for evidence retrieval [9]. All

extracted sections are split into short, overlapping spans (2k characters with

20% overlap). Each span is embedded with a dense encoder and stored in a

reusable vector index [12]. The rationale for this approach is twofold: criterion-

level retrieval is cheaper and typically more stable than passing the entire paper

to an llm; chunking supports evidence localization and mitigates failures where

information is missed or diluted in long inputs [14].

Paper Analysis. This branch implements the “paper-side” reproducibility check-

list. Following [18], it evaluates 20 criteria grouped into three categories: models,

datasets, and experiments. Each criterion has a precomputed embedding built

from its name, detailed description, and examples. At runtime, for criterion

embedding u and a section chunk embedding v, we compute cosine similarity

sim(u, v) =

u·v

∥u∥∥v ∥

and retrieve the top-k most similar chunks, with a conser-

vative similarity threshold and a token budget scheme. An llm then evaluates

one criterion at a time using only (a) the criterion deﬁnition and (b) the re-

trieved text spans, returning a structured record: presence p ∈ {0, 1}, conﬁdence

c ∈ [0, 1], and quoted evidence/notes (i.e., what makes the llm give that score

with the corresponding conﬁdence). Category scores are computed determin-

istically as S

= 100 ·

i∈I

), where p

and c

are respectively the

present/absent decision and conﬁdence for criterion i, and I

is the set of crite-

ria composing the category g. Finally, we combine category scores using paper-

type-speciﬁc weights to obtain a paper-side score S

paper

. Criterion-ﬁrst evalua-

tion yields interpretable failure modes (“missing split description” vs. “missing

license”), enables per-criterion auditing, and keeps scoring stable: llms extract

evidence; arithmetic is deterministic. The conﬁdence is used to tune the score,

assuming that lower values indicate a less clear presence.

Code Availability. This node gates the code branch. From the manuscript, we

resolve repository links and perform lightweight sanity checks (reachable, non-

empty, not a placeholder). If the code link is not available in the pdf, an online

search might be performed depending on conference policies. Both operations,

sanity check and online search, are supported by llms. If code is unavailable, the

workﬂow marks the code branch as skipped and proceeds without it. Fig. 1a is

generated from this node output.

6 F. Bolelli, D. Santoli, et al.

Code Embeddings. When code is available, a searchable repository index is

built in two passes to stay within a ﬁxed token budget while preserving repro-

ducibility-critical content. First, we ingest only the repository readme(s) plus

the full repository tree structure, including ﬁle sizes. We then ask an llm to

output an ordered list of include patterns (budgeted, e.g., ≤100k tokens to-

tal), prioritizing documentation, dependency specs (e.g., requirements.txt,

environment.yml), and training/evaluation entry points, while excluding pe-

ripheral scripts (visualizations, extra baselines, benchmarks). Finally, we re-

ingest only the selected ﬁles:

each ﬁle is chunked at word boundaries into

non-overlapping ∼20k-character blocks and embeddings are computed. Chunks

are stored with explicit positional metadata, i.e., ﬁle_path, chunk_index, to-

tal_chunks, to support evidence linking and retrieval-style lookups in down-

stream analysis. For integrity and cache invalidation, we compute and persist a

per-ﬁle SHA-256 digest so reruns can detect code changes. Ingestion is guided by

GitIngest [4]. We acknowledge that llm-driven ﬁle selection may be fragile, and

recommend venue-level guidelines that mandate clear, standardized repository

entry points.

Code Analysis. During this step, the repository snapshot, the embedded code

chunks, and paper spans produced by previous nodes are consumed. The analysis

is decomposed into six components (a superset of miccai’s code-oriented expec-

tations [18]): (i) research methodology detection (e.g., deep learning, machine

learning, algorithm, simulation, data analysis) together with applicability ﬂags

(e.g., requires training, requires datasets); (ii) repository structure and depen-

dency speciﬁcation quality; (iii) code components (e.g., training code, evaluation

code, documented commands); (iv) released artifacts (e.g., checkpoints, dataset

links, and coverage estimates); (v) dataset split speciﬁcation and random seeds;

(vi) documentation (e.g., readme, reproduction commands).

Each component is analyzed by a separate, targeted llm call with structured

output. Critically, the llm input is jointly grounded in repo evidence and paper

claims: for each component we assemble (a) the repository summary and ﬁle

tree, (b) the most relevant embedded code/documentation chunks (taken from

Code Embeddings), and (c) the corresponding paper spans (taken from Section

Embedding), so the model can leverage both data sources and ﬂag mismatches

between manuscript and repository rather than implicitly assuming consistency.

Both (b) and (c) are retrieved by cosine similarity against a component-speciﬁc

query, similarly to what is described in the Paper Analysis node.

Numeric scoring is fully deterministic. Given component outputs, we compute a

raw breakdown over (ii)—(vi) scoring axes with methodology-adaptive maxima

(e.g., max code-completeness is 3.0 if training is required, else 2.5; artifacts

are weighted up to 2.5 when datasets/training are required, else 2.0; splits are

weighted up to 2.0 only when splits are required). The overall score is then

normalized as S

code

= 100·

, where b

is the achieved score on axis j

The system is conﬁgured to always include critical ﬁlenames (e.g., eval.*, train.*).

Components (ii)–(vi) are methodology-dependent: (i) determines applicability, ex-

pected evidence, and thus the relative importance of each component in scoring.

Reproducibility Signals for Review Support 7

and m

its maximum score under the detected methodology; per-axis normalized

scores 100 · b

are also reported in output. Finally, recommendations are

generated programmatically from missing ﬂags (e.g., missing checkpoints when

requires training is true), ensuring actionable feedback without relying on the

llm for arithmetic or policy logic.

Dataset Analysis. If the classiﬁcation indicates a dataset contribution (e.g.,

the paper explicitly claims a new dataset), we run a dedicated dataset documen-

tation checklist (10 criteria derived from [18]) using the same retrieval-based,

structured-evaluation pattern as Paper Analysis, yielding S

data

. Indeed, dataset

papers have additional reproducibility obligations (collection protocol, annota-

tion, access restrictions, ethical statements) that are orthogonal to “method”

reproducibility and should be assessed explicitly.

Final Aggregation. Although intermediate scores are always reported, the

ﬁnal report merges criterion-level outputs from Paper Analysis, Code Analysis (if

available), and Dataset Analysis (if applicable) into a single “evaluation details”

artifact, stored for audit and human comparison. The overall score is a weighted

average with paper-type-dependent weights (e.g., methodological papers might

use S

final

= 0.6S

paper

+0.4S

code

). Such weights should reﬂect a pragmatic review

rationale and should be adapted to each conference.

During the aggregation stage, llm is used only to generate a short qualita-

tive synthesis (strengths/weaknesses) from already-computed results; it does not

generate or adjust numeric scores. The ﬁnal output includes overall and com-

ponent scores, a concise narrative summary, a machine-readable json artifact

with criterion-level decisions and evidence, and an interactive paper view with

highlighted supporting spans (Fig. 1b).

3.3 Infrastructure

Operationally, paper-snitch is deployed as a containerized web service via

Docker, so the full pipeline (web ui, workﬂow workers, parsers, and databases)

can be brought up reproducibly with pinned images and isolated dependencies.

An nginx reverse proxy serves static assets, forwarding api traﬃc to a Django

application. Long-running analyses are executed by a pool of Celery workers.

The analysis logic itself is expressed as a versioned graph in LangGraph (State-

Graph) [11], compiled and executed inside Celery tasks.

4 Performance in the Field

We evaluate paper-snitch on N=100 randomly sampled papers from the mic-

cai 2025 proceedings, stratiﬁed over the three main categories identiﬁed in the

Paper Classiﬁcation node (Sec. 3.2): dataset, method, and both.

For each paper, we run the full workﬂow twice, swapping only the llm used

in all generation steps: GPT-4o [7] and GPT-5 [22]. Embedding models, i.e.,

text-embedding-3-small with a ﬁxed embedding size of 1 536, and determin-

istic scoring code are kept ﬁxed across runs to isolate the eﬀect of the llm.

8 F. Bolelli, D. Santoli, et al.

Table 1: Human-machine agreement and I/O k-token usage per analysis axis.

Analysis

Category

GPT-4o GPT-5

% Agr. MCC

#IN #OUT

% Agr. MCC

#IN #OUT

Code 90.70 0.81 95.37 1.33 84.80 0.68 126.75 2.65

Paper 77.54 0.57 37.91 2.81 66.28 0.38 37.91 5.78

Dataset 72.22 0.29 13.39 0.99 65.00 0.18 13.39 1.71

Global 80.40 0.60 78.80 3.51 70.89 0.43 78.80 7.98

Whenever conﬁgurable, structured-output calls use low temperatures (0.1–0.2)

for near-deterministic decoding in criterion-level analyses, while synthesis tasks

use slightly higher values (0.3) to improve readability. Additionally, given the

miccai paper length, we set the retrieval depth to k=3 to ground the llm deci-

sion to a concise shortlist of evidence spans.

Human Reference Scores. We compare automated scores with human judg-

ments from two annotators with more than 3 years of experience. Annotators

ﬁlled the same criterion/component rubrics exposed by the tool (20 paper crite-

ria, 10 dataset criteria, and the code-component checklist) and evaluated sepa-

rate paper batches, discussing ambiguous cases to reach consensus before ﬁnal-

izing the rubric scores. In Tab. 1, the agreement between automated and human

labels is computed by treating each ﬁeld as a binary decision (true/false). Agree-

ment is computed independently for each of the three main analysis components

(paper, dataset, and code) and also overall (pooling all ﬁelds across components):

we report percent agreement (accuracy), deﬁned as the fraction of ﬁelds where

the automated label matches the human label, and Matthews Correlation Co-

eﬃcient (mcc) [17]. In our randomly sampled pool of 100 papers, paper-type

classiﬁcation was always correct. Tab. 1 also reports average token usage. The

cost per paper averages $0.23 for GPT-4o and $0.43 for GPT-5.

5 Discussion and Conclusion

This paper presents paper-snitch as a practical proof-of-concept for evidence-

grounded reproducibility screening in medical imaging submissions. The sys-

tem decomposes reproducibility into interpretable criteria, ties decisions to in-

spectable evidence, and produces auditable artifacts that can support human

judgment. We emphasize that weighting choices, criteria, and thresholds should

be tuned to each venue’s policies and validated before high-stakes deployment.

Limitations. Our evaluation uses a simple binary scheme, which may not cap-

ture partial releases, restricted-access data, or underspeciﬁed settings; future

work should evaluate richer schemes against expert judgments and reproducibil-

ity outcomes. The system also ingests untrusted text (e.g., readme) and re-

lies on llm-guided indexing, making it susceptible to prompt-injection; mitiga-

tions already include schema-constrained outputs and deterministic cross-checks

Reproducibility Signals for Review Support 9

against the repository tree. Finally, while llm costs are non-trivial, they ap-

pear feasible at conference scale and can be reduced via batching/caching or

on-premises/open-weight models when privacy or budget requires it.

Acknowledgments. This project has received funding from Fondazione di Mod-

ena, through the FAR 2024 (E93C24002080007), and from MUR, under the NRRP

“Fit4MedRob-Fit for Medical Robotics” (PNC0000007).

Disclosure of Interests. The authors have no competing interests to declare that

are relevant to the content of this article.

References

1. Association for Computing Machinery: Artifact Review and Badging. https://

www.acm.org/publications/policies/artifact-review-and-badging-current

(2020), accessed 2026-05-25

2. Chen, J., Qi, F., Chang, C., Hu, Q., Fu, K., Wang, X., Liu, K.: GLM-SFNet: Global-

Local Vision-Mamba with Semantic Fusion for Medical Image Segmentation. In:

Proceedings of Medical Image Computing and Computer Assisted Intervention –

MICCAI 2025 (2025)

3. Dai, P., Ou, Y., Yang, Y., Jin, Z., Suzuki, K.: GoCa: Trustworthy Multi-Modal

RAG with Explicit Thinking Distillation for Reliable Decision-Making in Med-

LVLMs. In: Proceedings of Medical Image Computing and Computer Assisted

Intervention – MICCAI 2025 (2025)

4. GitIngest: Turn any Git Repository into a Prompt-friendly Text Digest (2026),

https://gitingest.com/, accessed 2026-05-25

5. Gundersen, O.E., Kjensmo, S.: State of the Art: Reproducibility in Artiﬁcial Intel-

ligence. In: Proceedings of the AAAI Conference on Artiﬁcial Intelligence (2018)

6. Hoyt, R., Limon, A., Chang, A.: Generative AI and scientiﬁc manuscript peer

review. Intelligence-Based Medicine 11 (2025)

7. Hurst, A., et al.: GPT-4o System Card. arXiv preprint arXiv:2410.21276 (2024)

8. Jung, S., Lee, D., Kim, W.H.: MindLink: Subject-agnostic Cross-Subject Brain

Decoding Framework. In: Proceedings of Medical Image Computing and Computer

Assisted Intervention – MICCAI 2025 (2025)

9. Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., Yih,

W.t.: Dense Passage Retrieval for Open-domain Question Answering. In: Confer-

ence on Empirical Methods in Natural Language Processing (2020)

10. Kerautret, B., Bolelli, F., Colom, M., Lopresti, D. (eds.): Reproducible Research

in Pattern Recognition, Lecture Notes in Computer Science, vol. 15705. Springer

Cham (2025)

11. LangChain: LangGraph – Build Resilient Language Agents as Graphs (2026),

https://github.com/langchain-ai/langgraph, accessed 2026-05-25

12. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler,

H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-Augmented Generation

for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing

Systems (2020)

13. Liu, J., Li, H., Yang, C., Deutges, M., Sadaﬁ, A., You, X., Breininger, K., Navab, N.,

Schüﬄer, P.J.: HASD: Hierarchical Adaption for Pathology Slide-Level Domain-

Shift. In: Proceedings of Medical Image Computing and Computer Assisted Inter-

vention – MICCAI 2025 (2025)

10 F. Bolelli, D. Santoli, et al.

14. Liu, N.F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., Liang,

P.: Lost in the Middle: How Language Models Use Long Contexts. Transactions of

the Association for Computational Linguistics (2024)

15. Lopez, P.: GROBID: Combining Automatic Bibliographic Data Recognition and

Term Extraction For Scholarship Publications. In: International Conference on

Theory and Practice of Digital Libraries (2009)

16. Maier-Hein, L., Reinke, A., Kozubek, M., Martel, A.L., Arbel, T., Eisenmann, M.,

Hanbury, A., Jannin, P., Müller, H., Onogur, S., et al.: BIAS: Transparent reporting

of biomedical image analysis challenges. Medical Image Analysis 66 (2020)

17. Matthews, B.W.: Comparison of the predicted and observed secondary structure

of T4 phage lysozyme. Biochimica et Biophysica Acta (BBA)-Protein Structure

405(2) (1975)

18. MICCAI: Paper Submission Guidelines for Reproducible Research (2025),

https://conferences.miccai.org/2025/en/PAPER-SUBMISSION-GUIDELINES.

html#reproducibleresearch, accessed 2026-02-06

19. Naddaf, M.: AI is transforming peer review—and many scientists are worried.

Nature 639(8056) (2025)

20. NeurIPS: Paper Checklist Guidelines (2025), https://neurips.cc/public/

guides/PaperChecklist, accessed 2026-05-25

21. Pineau, J., Vincent-Lamarre, P., Sinha, K., Larivière, V., Beygelzimer, A., d’Alché

Buc, F., Fox, E., Larochelle, H.: Improving Reproducibility in Machine Learning

Research (A Report from the NeurIPS 2019 Reproducibility Program). Journal of

Machine Learning Research 22(164) (2021)

22. Singh, A., et al.: OpenAI GPT-5 System Card. arXiv preprint arXiv:2601.03267

(2026)

23. Stodden, V.C.: Reproducible Research: Addressing the Need for Data and Code

Sharing in Computational Science. Computing in Science & Engineering 12(5)

(2010)

24. Zhu, Z., Wang, J., Jiang, Y., Han, T., Huang, Y., Zhang, A., Yang, K., Luo, M.,

Liu, Z., Duan, Y., Ni, D., Tang, T., Yang, X.: Hierarchical Corpus-View-Category

Reﬁnement for Carotid Plaque Risk Grading in Ultrasound. In: Proceedings of

Medical Image Computing and Computer Assisted Intervention – MICCAI 2025

(2025)