RESEARCH PROJECT · ALON GORENSHTEIN, MD
LitBench: benchmarking evidence synthesis in epilepsy
LitBench evaluates whether AI systems can find the right papers, identify supporting passages, combine evidence across studies, and recognize when the evidence needed to answer is absent. The benchmark focuses on epilepsy research.
Research question
Can an AI research assistant find and synthesize the right evidence across a challenging medical literature corpus?
Approach and evidence
The medRxiv preprint describes 1,980 open-access papers and 9,472 human-reviewed facts, with single-paper and two-paper questions and tests where the supporting source is deliberately removed.
The authors report that no system was reliable across every tested dimension. More competing papers and cross-paper synthesis created challenges, and systems differed in their willingness to abstain when evidence was missing.
Authors and contributions
Alon Gorenshtein is first author, with Tianyu Zhang, M. Brandon Westover, and Daniel M. Goldenholz. The code is maintained in the Goldenholz Lab repository.
Research context
The linked 2026 manuscript is a preprint and has not been peer reviewed. Results reflect the models and conditions described in that manuscript.
Publications and resources
Research summary updated October 27, 2026.