Researchers present AutoResearch, a two-stage autonomous research system combining idea generation with idea execution. In the generation phase, the system integrates research signals with domain knowledge and uses multi-model generation to create testable plans; during execution, coordinated agents decompose plans into experiments, implement them iteratively, and apply evidence-based review before accepting conclusions. Evaluated across cross-modal retrieval, systems optimization, and machine learning benchmarks, it achieves a mean Recall improvement from 32.84 to 34.69 on RSICD while recording significantly fewer audit-confirmed issues than competing autonomous research systems.
