Auxiliary uncertainty signals for LLM-assisted systematic review screening: a benchmark across eight Cohen drug-class reviews
Not provided in the abstract
Abstract
This paper demonstrates that an auxiliary BERT+GCN classifier can improve the efficiency of LLM-assisted title-abstract screening in systematic reviews by providing structured uncertainty signals.
Reality Card
The study identifies that full-context delivery of prompts significantly improves screening efficiency metrics while MAYBE-only routing is found to be Pareto-optimal, achieving high recall at a lower cost.
Full-context delivery yields significant gains in F1 (+0.011) and WSS@95 (+0.050) at a 1.28x token-cost premium.
The study indicates that current instruction-tuned LLMs cannot self-triage, which may limit their adaptability in dynamic screening environments.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.