AUTOSYNTHESIS: An agentic system for automated meta-analysis

Evidence synthesis is crucial for turning primary research into reliable knowledge, yet quantitative evidence synthesis remains largely manual and difficult to scale. We introduce AUTOSYNTHESIS, an end-to-end multi-agent system for automated meta-analysis. Given a research question in natural language, AUTOSYNTHESIS formulates a search strategy, retrieves and screens scientific literature, assesses full-text eligibility, extracts and validates quantitative statistics, computes standardized effect sizes, and performs random-effects meta-analysis. It further supports heterogeneity analysis, publication-bias diagnostics, risk-of-bias assessment, and PRISMA-aligned reporting. We evaluate AUTOSYNTHESIS against an expert-conducted meta-analysis of LLM persuasion. AUTOSYNTHESIS retrieves 28 candidate records and synthesizes 20 effect-size estimates from 8 studies. It recovers 71.4\% of the benchmark studies, increasing to 85.7\% when the study missed solely because its full text was inaccessible is treated as recovered, and its pooled estimate differs from the human benchmark by $\Delta g = 0.123$ Hedges' $g$. These results demonstrate the feasibility of agentic systems for scalable, auditable quantitative evidence synthesis and point toward continuously updated ``living'' meta-analyses.