| name | agentslr-automated-literature-reviews |
| title | AgentSLR: Automating Systematic Literature Reviews |
| version | 0.0.3 |
| engine | skillxiv-v0.0.3-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.22327 |
| keywords | ["Systematic Literature Reviews","Epidemiology","Agentic AI","Evidence Synthesis","Domain Automation"] |
| description | Automate systematic literature reviews in epidemiology using agentic AI pipelines. Achieves 58x speed-up (7 weeks to 20 hours) by automating article retrieval, screening, data extraction, and report synthesis. Demonstrates that review quality depends on model capabilities rather than scale. Use when conducting evidence-based reviews in specialized domains, need to validate against human expertise, or require cost-effective evidence synthesis at scale. |
| category | Application Transfer |
Domain Problem
Systematic literature reviews (SLRs) in epidemiology are critical for evidence-based policy but present a severe bottleneck: the manual review process takes approximately 7 weeks of expert time per project. This limits how frequently health authorities can synthesize emerging evidence on priority pathogens, vaccines, and interventions. WHO and health agencies need rapid, reliable evidence synthesis to inform policy decisions during disease outbreaks and healthcare planning.
The Gap
Traditional SLR workflows require expert epidemiologists to manually screen articles, extract data, and synthesize findings. This is slow, expensive, and limits the frequency of evidence updates. Naive application of general-purpose LLMs fails because:
- Epidemiological screening requires domain-specific judgment about study design quality and relevance
- Data extraction from heterogeneous research papers demands understanding of study protocols, statistical methods, and outcome definitions
- Report synthesis must integrate evidence from disparate sources into actionable policy recommendations
Simply prompting a language model to "find relevant papers" produces high false-positive rates and misses domain-specific inclusion criteria.
Source Technique
The foundation is multi-step agentic AI workflows where an LLM orchestrates a pipeline of specialized tasks. Rather than a single forward pass, the agent iterates through retrieve-screen-extract-synthesize steps, maintaining context and refining decisions based on intermediate results. This mimics the human SLR process: humans don't decide inclusion once; they refine criteria as they encounter papers.
Adaptation Recipe
AgentSLR bridges the gap between general LLM capability and epidemiological SLR requirements through:
-
Workflow Decomposition: Break the monolithic "conduct a literature review" task into substeps (retrieve articles, screen by title/abstract, screen by full text, extract structured data, synthesize findings). Each step can be validated independently against expert judgment.
-
Domain-Specific Prompting: Provide the LLM with explicit inclusion/exclusion criteria, outcome definitions, and quality metrics from the research protocol before executing retrieval. This anchors the agent to domain constraints rather than letting it invent criteria.
-
Human-in-the-Loop Validation: Identify failure modes through comparison with expert-curated benchmarks. Rather than trusting the agent fully, use expert validation to detect when the agent diverges from domain norms (e.g., missing important study designs or applying criteria inconsistently).
-
Multi-Model Capability Matching: Test across frontier models and select based on task-specific performance, not raw scale. The paper found that "performance is driven less by model size or inference cost than by each model's distinctive capabilities." Some models excel at screening, others at data extraction—match the model to the task.
