AutoResearch grounds autonomous research by counting audit-confirmed failures, not just outcomes

arXiv·low signalXBlueskyLinkedInCopy link

A two-stage system connects idea generation, which fuses emerging research signals with accumulated domain knowledge through multi-model generation and cross-review, to idea execution, where coordinated agents decompose plans into experiments and run independent evidence-based review before accepting any conclusion. On the RSICD cross-modal retrieval benchmark an AutoResearch-generated idea lifts mean Recall from 32.84 to 34.69 while logging only 5 audit-confirmed issue events against 11 to 27 for other autonomous research systems. The audit-event count is the transferable idea: it makes process reliability a reportable metric rather than an assumption.