01 / the method
About FLASH
A living research framework for accelerating literature screening in systematic reviews of randomized clinical trials
What is FLASH?
A framework built for evidence synthesis
FLASH stands for Framework for LLM-based Assisted Screening with High sensitivity
It is a living framework that uses large language models to accelerate literature screening for systematic reviews of randomized clinical trials
The method standardizes eligibility criteria and applies screening prompts designed to follow the reasoning used in manual screening
It supports both abstract and full-text screening
FLASH was developed and validated on FLASHBench, a large-scale, representative and living benchmark for literature screening
The research method
Criteria first
Screening in stages
The research method is the foundation for the online service now in development
- 01
Standardize eligibility criteria
Translate the review's eligibility criteria into a consistent form before screening begins
- 02
Screen each record independently
The research configuration uses three language models: GPT-5.5 (OpenAI), Qwen3.8-Max (Alibaba Cloud) and DeepSeek-V4-Pro (DeepSeek)
Each model screens independently; records included by any one model are retained - 03
Screen abstracts and full texts
Apply the criteria at the abstract stage, followed by full-text screening of the retained records
This three-model configuration describes the research method
The first online Screening Service is planned with a separate two-model configuration
The research results below do not establish validation of that online configuration
Research findings
Reducing the screening workload
More than 92% reduction in screening workload
This is a research workload measure, not an accuracy measureThe research validation reports 99.9% sensitivity for eligible randomized clinical trials identified through manual screening
FLASH also detects eligible trials missed by manual screening
These findings describe the research validation of FLASH
They are not a performance guarantee for the forthcoming online Screening Service