Adversarial Multi-Agent System for Systematic Literature Reviews: Author–Reviewer Workflows with Verifiable Evidence and Critique Loops

1. Overview

1.1 Scientific rationale and importance

Systematic literature reviews (SLRs) underpin evidence-based biomedical research by rigorously synthesising all relevant studies to reach reliable conclusions, while also informing robust experimental design and biomarker selection and reducing selective interpretation [1]. However, SLRs are time- and labour-intensive [2]. A recent protocol-tracking study reported a median of 11.5 months to complete and 16.2 months to publish an SLR [3]. Although the emergence of large language models has attracted some researchers to explore automated workflows to accelerate the completion of systematic literature reviews (SLRs) [4], existing workflows lack rigorous auditability (i.e., verifiable extraction trails) and remain highly susceptible to bias (e.g., query/coverage and selection bias), limiting the reliability of their synthesised conclusions. To address these limitations, this project will build and evaluate an automated adversarial multi-agent author–reviewer workflow for conducting SLRs. The “author” agents execute full PRISMA-protocol SLR steps (protocol drafting → query design → multi-source retrieval → normalisation → deduplication → screening → structured data extraction → synthesise figures, tables, and reports), while auditable evidence ledgers (claim → source → locator/snippet) constrain generation to verifiable statements and improve auditability. To avoid SLR bias, independent “reviewer” agents provide comments on completeness, bias, and evidence support. Through the competition between the two, the quality of the generated SLR will continuously improve.

1.2 Confirmation of funding/facilities/ethics approvals

Funding: This proposal requests support under the ResearchHub RFP “AI Tools & AI Agents for Biomedical Research” (USD $5,000) to cover model/API usage, engineering, and expert evaluation.

Facilities: Development will be conducted on standard institutional computing resources (personal workstation and/or institutional servers) with internet access for literature APIs, plus version-controlled code hosting (e.g., GitHub).

Ethics approvals: Ethics approvals for data handling and workflow evaluation will be obtained. As the project does not involve human participants, patient data, or animal experiments, the ethical risk is low. It operates on public bibliographic metadata and open-access literature where available (e.g., PubMed/PMC, Europe PMC). The evaluation process will be collected as non-sensitive, minimal-risk input (no personal data beyond optional role/title), and handled under standard institutional guidelines; participation will be opt-in and anonymized in shared outputs.

1.3 Team & Estimated Timeline

This project will be conducted by a team with complementary skills, including Dr Susan He and Dr Zhipeng Yuan. Dr Susan He is a postdoctoral research fellow at the Institute for Regeneration and Repair, University of Edinburgh. She has extensive experience in cancer and cellular senescence and will lead the project, overseeing experimental validation and reporting of the workflow. Dr Zhipeng Yuan, Postdoctoral research associate in the School of Computer Science at the University of Sheffield, UK, has extensive experience in multi-agent construction. His research interests include responsible AI and the application of LLM in industry and academia. In this project, he will be responsible for building the SLR workflow through Python and the A2A protocol.

The estimated timeline for this project is shown below.

· Stage 1 (Months 1–2): “Author” agent construction — Build the protocol-first author agent SLR pipeline (retrieval → dedup → screening → extraction → evidence-led synthesis) to generate draft manuscripts plus PRISMA-style figures and summary tables.

· Stage 2 (Months 2–3): “Reviewer” agent construction — Add reviewer agents and complete the critique–revise loop, enabling automatic revision/regeneration of manuscripts, tables, and figures based on auditability and bias-focused reviewer comments.

· Stage 3 (Month 3-5): Evaluation in two biomedical case studies — Evaluate on two biomedical case studies with a 2×2 design (Baseline vs Adversarial; ChatGPT vs DeepSeek), using quantitative metrics and expert feedback.

· Stage 4 (Month 5-6): Reporting and dissemination — Release the open-source workflow and two runnable tutorials, write a preprint academic paper to support future publications, and publish dissemination resources (manual, videos, GitHub Pages).

1.4 Data Sharing Commitment

We commit to open and reproducible sharing of outputs. Specifically:

· Open-source code (pipeline, agent prompts/configs, schemas, tests) under a permissive license.

· Reproducibility package for each demo run: saved queries, timestamps, retrieval counts, dedup traces, screening decisions, extraction tables, reviewer critiques, and revision logs.

· We will not redistribute copyrighted full-text content; instead, we will share identifiers (DOIs/PMIDs/PMCIDs), metadata, and minimal excerpts/locators when permitted. Any human feedback used for evaluation will be anonymized before sharing.

2. Introduction

2.1 Background

Systematic literature reviews (SLRs) underpin evidence-based biomedical research by transparently identifying, appraising, and synthesising all relevant studies to support reliable conclusions [1]. However, conducting SLRs is time- and labour-intensive, which makes it difficult to rigorously reproduce and validate on a large scale [2]. With the emergence of large language models (LLMs) in recent years, their applications demonstrate the ability to accelerate the completion of SLRs with text processing and understanding capabilities [4].

2.2 Related Works

Recent work has outlined how LLMs might be integrated into systematic review workflows by automated information extraction [5] and how RAG/agentic orchestration may reduce hallucination by grounding outputs in literature for synthesising scientific literature [6]. In addition, LLMs are also being explored for end-to-end scientific manuscript drafting (e.g., structured section generation [7], citation-aware writing [8]) and for automating aspects of peer review, such as critique generation [9], methodological checking [10], and meta-review style decision support [11], which demonstrates the ability of LLMs for academic writing. Table I summarises the recent related works. These studies suggest that LLMs can substantially accelerate SLRs from a full process. Although LLMs have motivated automated SLR workflows, most existing studies do not implement a full-process pipeline and struggle with two core limitations: (i) auditability (lack of claim-level provenance and verifiable extraction trails) and (ii) reliability under bias (insufficient mechanisms to detect/query drift, coverage gaps, and selective reporting; limited robustness checks).

Related work	Category	Main contribution	Limitations
PRISMA 2020 [12]	Standards & governance	Standard checklist + flow diagram for transparent SR reporting	Reporting standard only; does not enforce claim-level provenance or bias-robust execution
Cochrane Handbook [13]	Standards & governance	Authoritative protocol-driven methods for trustworthy evidence synthesis	Manual compliance is costly; no built-in automation for provenance auditing or debias stress-tests
Prognostic factor SR guide [1]	Standards & governance	Methods for systematic review/meta-analysis of prognostic factors (biomarker-relevant)	Guidance only; does not provide scalable audit trails or bias/coverage checks
SR automation tools scoping review [14]	Traditional SR automation tools	Maps SR automation tools across workflow stages	Tool fragmentation; limited end-to-end provenance and limited bias/robustness verification
Rayyan [15]	Traditional SR automation tools	Popular human-in-the-loop screening platform	Primarily screening; no manuscript-level evidence ledger or adversarial debias auditing
ASReview LAB v2 [16]	Traditional SR automation tools	Active-learning screening to reduce workload	Screening efficiency focus; weak on claim→evidence auditability and coverage/temporal bias checks
RobotReviewer [17]	Traditional SR automation tools	Automates risk-of-bias judgments with supporting text spans	Task-specific; not end-to-end SLR generation with full provenance + debias checks
LLMs in SR framework [5]	LLM-assisted evidence synthesis & writing	Framework/case study for integrating LLMs into SR tasks	Mostly guidance; lacks enforced claim-level provenance and adversarial reviewer loop by default
LLM review generation [8]	LLM-assisted evidence synthesis & writing	Demonstrates LLM-driven review generation with QC ideas	Not fully SLR/PRISMA protocolized; limited explicit debias stress-testing and independent auditing
RAG scientific synthesis [6]	LLM-assisted evidence synthesis & writing	Citation/passage-grounded scientific synthesis via RAG	Not a full SLR pipeline (screening/extraction standards); no author–reviewer adversarial loop
LLM CONSORT auditing [10]	LLM-assisted evidence synthesis & writing	Automated methodological/reporting compliance checking at scale	Audits reporting; doesn’t address retrieval/selection bias or provide full claim-level SLR provenance
LLM paper feedback [9]	Multi-agent critique & verification	Large-scale evidence LLMs can provide useful critique	Not integrated with SLR artifacts (ledger/PRISMA/tables); critique not guaranteed to be evidence-grounded
Agent-based peer review [11]	Multi-agent critique & verification	Role-based author/reviewer/meta-reviewer agents for critique loops	Not SLR-specific; needs adaptation to enforce provenance completeness and bias/robustness checks

2.3 Research Questions

For the shortcomings of existing LLM-assisted SLR workflows, we focus on proposing an adversarial multi-agent “author-reviewer” workflow for a full-process SLR generation, which will answer the following research questions (RQs) in this project.

· RQ1 (Auditability): Can an “author” workflow increase end-to-end auditability by producing a machine-readable evidence ledger that links each extracted datum and manuscript claim to specific source locators?

· RQ2 (Reliability / debias): Can independent “reviewer/red-team” agents measurably reduce bias-driven errors in SLRs (such as query/coverage and selection bias, selective reporting, temporal staleness) compared with single-pass automated SLR generation?

· RQ3 (Efficiency with safeguards): Can the workflow reduce time-to-search and time-to-draft (tables/figures/manuscript) in conducting SLRs while maintaining or improving reliability metrics (unsupported-claim rate, coverage gaps, protocol compliance)?

2.4 Hypotheses

This project will validate whether the adversarial multi-agent SLR workflow, where “author” agents generate protocol, retrieval, screening, extraction, figures/tables, and manuscript drafts, and independent “reviewer/red-team” agents challenge evidence support and bias, can simultaneously improve auditability, reliability (debiasing), and efficiency.

· H1: Evidence-ledger enforcement will increase provenance coverage and reduce unsupported claims in the generated manuscript.

· H2: Reviewer/red-team agents will identify and correct more bias-related issues (missing key studies, protocol drift, selective emphasis, stale searches) than non-adversarial pipelines.

· H3: The author–reviewer loop will cut SLR production time substantially (target: reduce typical timelines from ~11–16 months to ~3–6 weeks for a “first-pass” SLR package, depending on topic size), while keeping auditability and reliability checks intact.

2.5 Importance, Innovation, and impact

SLRs commonly take months to over a year, e.g., PROSPERO-based analyses estimate a mean of ~67.3 weeks (~15.5 months) from start to publication, and protocol-tracking reports medians of ~11.5 months to complete and ~16.2 months to publish [3]. This massive workload hinders biomedical researchers from exploring new research fields comprehensively and occupies a large amount of work time that could have been used for more valuable experimental exploration. An automated and reliable SLR workflow can further assist biomedical researchers in focusing on valuable explorations and accelerate the development of the biomedical field. Although this is not the first project to utilise LLMs to assist SLR workflow, it will be the first full-process SLR workflow that emphasises auditability and reliability as far as we know.

This project’s impact is to make SLR production and updating dramatically more efficient, auditable, reliable, and reproducible by delivering an adversarial author–reviewer multi-agent workflow that automatically generates the core SLR package (protocol, screening/extraction logs, PRISMA-style figure, summary tables, and a manuscript draft with claim-level evidence links). We anticipate reducing the time to a first complete, auditable SLR draft package from months to weeks (target ~2–4 weeks per topic, depending on scope), and enabling rapid updates (re-running searches and revising tables/figures/manuscript) in hours to days rather than restarting the workflow. Regardless of outcome, the project will provide an open, reproducible document and code that shifts SLR effort from manual labour to automated agentic workflow, improving throughput for biological researchers while maintaining reliability and transparency.

2.6 Objectives

To construct and validate the proposed adversarial multi-agent workflow, this project have following objectives

· Objective 1 — Develop an adversarial “author–reviewer” multi-agent SLR workflow: implement “author” agents to automate the full SLR pipeline (protocol, query/retrieval, deduplication, screening, extraction, and generation of publication-ready tables/figures/manuscript), and “reviewer/red-team” agents to enforce auditability (claim-level evidence provenance, traceable extraction) and improve reliability/debiasing (detect query/coverage/selection bias and temporal staleness) through critique-and-revision loops.

· Objective 2 — Evaluate the system in two biomedical case studies by quantifying efficiency (time-to-first auditable SLR package and time required to update the review) and quality (provenance coverage, unsupported-claim rate, missing-study detection, and protocol compliance/robustness), alongside structured expert feedback. The two case studies will be: (1) Senescent cell burden and senolytic interventions in humans, and (2) Therapy-induced senescence in cancer: clinical associations and targetability. Outputs from these case studies are expected to provide the foundation for two publishable SLRs, following expert refinement and validation.

· Objective 3 — Disseminate reusable research artefacts: release open-source code/Python packages for the proposed agentic SLR workflow, step-by-step tutorials/demos (e.g., ResearchHub updates and a runnable example), and a preprint article to showcase project achievements and experience. The preprint article will be expanded and published in top academic conferences or journals in subsequent projects.

3. Methods

3.1 Technical approach (Agentic workflow design)

We will develop a role-separated, adversarial multi-agent workflow operating over a shared, versioned workspace (run ID, protocol version, query logs, canonical paper table, screening/extraction outputs, and an Evidence Ledger). “Author” agents will generate the SLR artifacts end-to-end, while “reviewer/red-team” agents will independently audit auditability (claim-level provenance) and reliability/debias (query/coverage/selection bias and temporal staleness). An arbiter/meta-reviewer agent will consolidate critiques into a revision checklist and will require iterative updates until pass criteria are met.

3.1.1 Author agents:

The “author” workflow will be implemented as a fixed, seven-step pipeline, with each step producing versioned artifacts that feed into the next stage:

1) Protocol Author
The agent will define the review protocol, including PICO/PECO framing, inclusion/exclusion criteria, primary/secondary outcomes, target databases/sources, date cutoffs, and planned analyses. The protocol will be stored as a versioned configuration to prevent drift.

2) Search & Retrieval
The agent will design and expand 1–3 seed queries (synonyms, MeSH-like variants, inclusion/exclusion constraints), will execute multi-source retrieval (e.g., PubMed/PMC, Europe PMC, Crossref/OpenAlex for enrichment), and will log all queries, timestamps, and retrieval counts/results for reproducibility.

3) Normalize & Deduplicate
The agent will canonicalize identifiers (prefer DOI; otherwise PMID/PMCID), will merge duplicates using deterministic rules plus fuzzy title+year matching when needed, and will generate a “dedup trace” documenting merge rationale and confidence.

4) Screening
The agent will perform two-stage screening (title/abstract → full text where available) using an explicit rubric aligned to the protocol. Each decision will include an include/exclude label, short rationale, confidence score, and the invoked protocol clause.

5) Extraction
The agent will extract structured study fields (e.g., study type, model/population, methods, endpoints/markers, key findings, limitations) and will attach evidence locators/snippets for each field to enable auditing.

6) Synthesis
The agent will generate an IMRaD-style manuscript draft (and related work/background) that is constrained by the evidence ledger, such that each claim is linked to one or more supporting sources/locators.

7) Artefact Builder
The agent will produce publication-ready outputs, including a PRISMA-style flow figure, structured summary tables (study characteristics/outcomes), and optional risk-of-bias/quality tables if enabled by the protocol.

All intermediate and final outputs will be saved as auditable appendices (covering protocol, logs, tables, figures, manuscript, and the evidence ledger) to support independent verification and rapid updates.

3.1.2 Reviewer Agents

We will build a small set of independent “reviewer” agents that audit the artifacts produced by the author pipeline (protocol, query logs, canonical paper table, dedup trace, screening decisions, extraction tables, manuscript draft, figures/tables, and the Evidence Ledger). Each reviewer will output structured findings (issue type, severity, evidence, required fix), and an arbiter/meta-reviewer will gate revisions until pass criteria are met.

· Evidence Integrity Reviewer (auditability): checks that every key claim and extracted field is linked to supporting sources/locators in the Evidence Ledger; flags unsupported claims and inconsistencies (e.g., numbers/effect direction).

· Coverage & Bias Reviewer (reliability/debias): probes query/coverage and selection bias by identifying missing key studies, inconsistent screening rationales, and protocol drift; recommends targeted query updates and re-screening.

· Manuscript Quality Reviewer (manuscript perspective): provide peer-review–style comments on clarity, structure (IMRaD coherence), completeness, and alignment between narrative, tables, and figures; it also run lightweight methodological/reporting checks (e.g., whether required reporting elements are present), similar in spirit to automated guideline-auditing approaches

An arbiter/meta-reviewer will consolidate reviewer feedback into a revision checklist and will interact with the author pipeline through a simple two-stage critique–revise cycle to improve SLR quality. After the reviewers critique the generated documents for evidence integrity, coverage/bias risks, and manuscript quality, and the arbiter will convert their findings into a prioritized revision checklist. The author agents will then revise the upstream steps that caused the issues (e.g., queries, screening decisions, extraction fields) and automatically propagate updates downstream so the manuscript, tables, and PRISMA outputs remain consistent, repeating this critique–revise cycle until quality thresholds are met.

3.2 Materials and Procedures

After implementation of the proposed workflow, we will evaluate the workflow on two biomedical SLR case studies, each defined by a protocol (PICO/PECO, inclusion/exclusion, outcomes, databases, and date cutoffs). The topics will be selected to differ in pace/complexity to stress-test coverage bias and temporal staleness.

3.2.1 Materials

Retrieval sources (expandable adapters). PubMed/PMC (NCBI Entrez), Europe PMC, Crossref, OpenAlex. Output: a canonical paper table (DOI/PMID/PMCID/title/authors/year/journal/abstract/links/source).

Base models. ChatGPT (commercial) and DeepSeek (open-source) will be used in complementary author/reviewer roles for cross-checking.

Core artifacts. Protocol config; query logs; dedup trace; screening JSON (decision/rationale/confidence); extraction tables with evidence locators/snippets; Evidence Ledger (claim→source→locator); manuscript draft; PRISMA-style flow figure and summary tables.

Devices / equipment and setting.

A primary development workstation (Windows/Linux/macOS) for orchestration, prompt/version control, and report generation.
Optional institutional compute/server (CPU/GPU as available) for running the open-source DeepSeek model locally and batch processing.
Secure storage (institutional drive or encrypted local storage) and a version-controlled repository (e.g., Git).
Reproducible runtime with pinned environments (container/lockfile), and run tracking (run_id, timestamps).

3.2.2 Procedures

For each of two biomedical case studies, we will evaluate both workflow condition and base model in a factorial design:

Workflow condition: Baseline (author-only, single-pass) vs Adversarial (author + reviewer critique–revise).
Base model: ChatGPT (commercial) vs DeepSeek (open-source).

This yields 4 runs per case study:
(1) Baseline+ChatGPT, (2) Baseline+DeepSeek, (3) Adversarial+ChatGPT, (4) Adversarial+DeepSeek.
All runs will follow the same fixed author pipeline: Protocol Author → Search & Retrieval → Normalize & Deduplicate → Screening → Extraction → Synthesis & Writing → Artifact Builder, producing protocol/logs/PRISMA figure/tables/manuscript/Evidence Ledger.

Human expert interaction (2–3 experts): Experts will review a standardized output pack for each run (manuscript + key tables/figures + a claim-evidence sample). They will label a sample of claims as supported/unsupported/unclear, spot-check a small subset of screening/extraction decisions, and provide brief Likert ratings (trust, clarity, ease of verification, time saved, quality) with short comments after reading the generated SLRs. Annotations will be collected in structured datasets.

3.2.3 Evaluation metrics and how they answer RQs / hypotheses

RQ1 / H1 (Auditability):

Provenance coverage (%): % claims linked to ≥1 Evidence Ledger entry.
Unsupported-claim rate (%): reviewer-flagged and expert-validated on a claim sample.
Evidence completeness (%): % extracted fields with evidence locators/snippets.

RQ2 / H2 (Reliability / debias):

Missing-study detection: additional key studies identified and incorporated after critique.
Protocol drift / inconsistency: decisions conflicting with protocol clauses.
Staleness/robustness: change in included-study set and conclusions after updated search cutoffs and small query perturbations (overlap/Jaccard; # conclusion changes).

RQ3 / H3 (Efficiency):

Time-to-first auditable SLR package and time-to-update.
Iteration cost: number of critique–revise cycles required.

Expert ratings for trust/clarity/ease-of-verification/time saved + qualitative feedback.

How we will analyse model vs workflow effects: We will compare (i) Baseline vs Adversarial within each model, and (ii) ChatGPT vs DeepSeek within each workflow condition, reporting main effects and interactions using paired summaries (absolute/relative differences; bootstrap confidence intervals where appropriate), per case study and aggregated across both case studies.

3.3 Expected Project Deliverables

During this project, we expect the project to output the following deliverables.

1) Open-source multi-agent SLR workflow
A released codebase/Python package implementing the adversarial author–reviewer SLR pipeline (protocol→retrieval→dedup→screening→extraction→tables/figures/manuscript) with Evidence Ledger auditability and debias/staleness checks.

2) Two case-study tutorials (runnable examples)
Step-by-step, reproducible tutorials for the two biomedical case studies, including configs, scripts/notebooks, and packaged outputs demonstrating baseline vs adversarial and ChatGPT vs DeepSeek runs.

3) A pre-print paper
An academic-style preprint paper (or structured report) describing the method, evaluation design, metrics, quantitative results, limitations, and recommended usage patterns.

4) Dissemination resources
User-facing materials to support adoption: user manual, short demo videos, and a simple project website (e.g., GitHub Pages) linking documentation, tutorials, releases, and example outputs.

3.4 Project Plan (6-month in total)

Stage 1 (Months 1–2): Author agent construction

Tasks:

Build protocol-first author agents: Protocol Author → Search & Retrieval → Normalize & Deduplicate → Screening → Extraction → Synthesis & Writing → Artifact Builder
Implement retrieval adapters (PubMed/PMC, Europe PMC, Crossref, OpenAlex) and generate the canonical paper table
Produce initial publication-ready outputs: IMRaD draft, PRISMA-style flow counts/figure, study summary tables (optional RoB template)

Deliverables: Author-agent workflow (v1) with retrieval adapters; canonical paper table schema; manuscript/table/figure templates.

Stage 2 (Months 2–3): Reviewer agents + author revision loop

Tasks:

Implement reviewer agents: Evidence Integrity, Coverage & Bias, Manuscript Quality
Build the Evidence Ledger + audit report outputs (claim→source→locator; extraction traceability)
Upgrade author agents to revise queries/screening/extraction and rewrite/regenerate manuscript, tables, and figures in response to reviewer comments
Implement the critique–revise loop with an arbiter/meta-reviewer, revision checklist, and release thresholds; integrate ChatGPT + DeepSeek role assignment/cross-checking

Deliverables: Completed end-to-end adversarial author–reviewer workflow (generation + critique + revision) with gating thresholds and reproducibility logging.

Stage 3 (Months 3–5): Evaluation in two biomedical case studies

Tasks:

Run the 2×2 evaluation: Baseline vs Adversarial × ChatGPT vs DeepSeek, for two topics
Measure: provenance coverage, unsupported-claim rate, missing-study detection, protocol compliance, robustness (query/date perturbations), and efficiency (time-to-first package, update time)
Collect lightweight expert feedback (claim checks + usability ratings + comments)

Deliverables: Two case-study reproducibility bundles (queries/logs/ledgers/tables/figures/manuscripts) as tutorials; evaluation metrics report + expert feedback summary.

Stage 4 (Months 5–6): Reporting and dissemination

Tasks:

Finalize results and write a preprint/manuscript-style report describing method, evaluation, and limitations
Release open-source code/Python package + runnable tutorials for both case studies
Prepare dissemination resources: user manual, short demo videos, and a GitHub Pages website; post updates on ResearchHub/community channels

Deliverables: Public repo/package + tutorials; preprint/result report; documentation + videos + project website.

3.5 Ethics and Data Management

1) Ethics status

No patient data, no interventions, no animal work. Primary data are public bibliographic records and OA texts where available. Formal IRB/ethics approval is typically not required.
Human feedback (3–5 experts) is minimal-risk; participation is voluntary with informed consent, and feedback is stored without direct identifiers.

2) Confidentiality measures

Do not store personal data beyond optional role/affiliation category (if provided).
Anonymise rater feedback before sharing (aggregate scores + de-identified comments).

3) Data sharing / archiving plan

Release: open-source code, prompts/configs, schemas, and reproducibility artifacts (queries, logs, dedup traces, screening/extraction tables, audit reports, figures, manuscript drafts).
Do not redistribute copyrighted full texts; share only IDs/metadata and permissible short snippets/locators.
Archive releases with version tags and a workflow/model card describing intended use, limitations, and evaluation results.

4. Budget(USD $5,000)

Compute / model usage / API costs: $1,000
Engineering time support (testing, packaging, documentation): $2,800
Lightweight expert review/annotation (small honoraria): $1,000
Contingency (rate limits, tooling, minor services): $200

5. References

[1]. Riley, R.D., Moons, K.G., Snell, K.I., Ensor, J., Hooft, L., Altman, D.G., Hayden, J., Collins, G.S. and Debray, T.P., 2019. A guide to systematic review and meta-analysis of prognostic factor studies. bmj, 364.

[2]. Borah, R., Brown, A.W., Capers, P.L. and Kaiser, K.A., 2017. Analysis of the time and workers needed to conduct systematic reviews of medical interventions using data from the PROSPERO registry. BMJ open, 7(2), p.e012545.

[3]. Bugajska, J.V., Hild, B.F., Brüschweiler, D., Meier, E.D., Bannach-Brown, A., Wever, K.E. and Ineichen, B.V., 2025. How long does it take to complete and publish a systematic review of animal studies?. BMC Medical Research Methodology, 25(1), p.226.

[4]. Lieberum, J.L., Toews, M., Metzendorf, M.I., Heilmeyer, F., Siemens, W., Haverkamp, C., Böhringer, D., Meerpohl, J.J. and Eisele-Metzger, A., 2025. Large language models for conducting systematic reviews: on the rise, but not yet ready for use—a scoping review. Journal of Clinical Epidemiology, 181, p.111746.

[5]. Hasan, B., Saadi, S., Rajjoub, N.S., Hegazi, M., Al-Kordi, M., Fleti, F., Farah, M., Riaz, I.B., Banerjee, I., Wang, Z. and Murad, M.H., 2024. Integrating large language models in systematic reviews: a framework and case study using ROBINS-I for risk of bias assessment. BMJ evidence-based medicine, 29(6), pp.394-398.

[6]. Asai, A., He, J., Shao, R. et al. Synthesizing scientific literature with retrieval-augmented language models. Nature (2026). https://doi.org/10.1038/s41586-025-10072-4

[7]. Kobak, D., González-Márquez, R., Horvát, E.Á. and Lause, J., 2025. Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27), p.eadt3813.

[8]. Wu, S., Ma, X., Luo, D., Li, L., Shi, X., Chang, X., Lin, X., Luo, R., Pei, C., Du, C. and Zhao, Z.J., 2025. Automated literature research and review-generation method based on large language models. National Science Review, 12(6), p.nwaf169.

[9]. Liang, W., Zhang, Y., Cao, H., Wang, B., Ding, D.Y., Yang, X., Vodrahalli, K., He, S., Smith, D.S., Yin, Y. and McFarland, D.A., 2024. Can large language models provide useful feedback on research papers? A large-scale empirical analysis. NEJM AI, 1(8), p.AIoa2400196.

[10]. Srinivasan, A., Berkowitz, J., Friedrich, N.A., Kivelson, S. and Tatonetti, N.P., 2025. Large language model analysis of reporting quality of randomized clinical trial articles: a systematic review. JAMA Network Open, 8(8), p.e2529418.

[11]. Jin, Y., Zhao, Q., Wang, Y., Chen, H., Zhu, K., Xiao, Y. and Wang, J., 2024, November. AgentReview: Exploring Peer Review Dynamics with LLM Agents. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 1208-1226).

[12]. Page, M.J., McKenzie, J.E., Bossuyt, P.M., Boutron, I., Hoffmann, T.C., Mulrow, C.D., Shamseer, L., Tetzlaff, J.M., Akl, E.A., Brennan, S.E. and Chou, R., 2021. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. bmj, 372.

[13]. Chandler, J., Cumpston, M., Li, T., Page, M.J. and Welch, V.J.H.W., 2019. Cochrane handbook for systematic reviews of interventions. Hoboken: Wiley, 4(1002), p.14651858.

[14]. Khalil, H., Ameen, D. and Zarnegar, A., 2022. Tools to support the automation of systematic reviews: a scoping review. Journal of Clinical Epidemiology, 144, pp.22-42.

[15]. Ouzzani, M., Hammady, H., Fedorowicz, Z. and Elmagarmid, A., 2016. Rayyan—a web and mobile app for systematic reviews. Systematic reviews, 5(1), p.210.

[16]. de Bruin, J., Lombaers, P., Kaandorp, C., Teijema, J.J., van der Kuil, T., Yazan, B., Dong, A. and van de Schoot, R., ASReview LAB v2: Open-Source Text Screening with Multiple Agents and Oracles. Available at SSRN 5136987.

[17]. Marshall, I.J., Kuiper, J. and Wallace, B.C., 2016. RobotReviewer: evaluation of a system for automatically assessing bias in clinical trials. Journal of the American Medical Informatics Association, 23(1), pp.193-201.

Adversarial Multi-Agent System for Systematic Literature Reviews: Author–Reviewer Workflows with Verifiable Evidence and Critique Loops

Funding Progress

Nonprofit Recipient

DOI

Adversarial Multi-Agent System for Systematic Literature Reviews: Author–Reviewer Workflows with Verifiable Evidence and Critique Loops

Funding Progress

Nonprofit Recipient

DOI