San Francisco – September 07, 2026 -- EvoMap has open-sourced AutoResearch, a multi-agent system that converts AI-generated research hypotheses into executable experiments and uses the resulting evidence to decide whether to continue, revise or abandon a research direction.
AutoResearch lifts a Django benchmark fix from 2/7 to 7/7 through iterative testing
On a real Django issue drawn from SWE-bench Lite, AutoResearch initially scored 2 out of 7 on official new-feature tests, then improved to 4 out of 7. Rather than halting at the partial gain, the system continued investigating the underlying problem and ultimately reached 7 out of 7, while holding 203 out of 203 regression tests intact.
System raises RSICD benchmark recall from 32.84 to 34.69 via AI-generated hypothesis
On the RSICD benchmark, a research idea generated by AutoResearch improved mean Recall from 32.84 to 34.69, demonstrating that an AI-proposed hypothesis could be converted into a measurable performance gain through repeated experimentation rather than a single untested claim.
Multiple AI models cross-review ideas before converting them into resourced experiment plans
AutoResearch uses several AI models to independently generate and cross-review research ideas. Accepted ideas are converted into executable research plans with defined metrics, success criteria, resource budgets and evaluation procedures, while specialized agents handle planning, implementation, experimentation, analysis and review. Research state, code, experiment logs, metrics, failures and prior decisions are kept in a persistent workspace so work can resume rather than restart, and independent blind review challenges conclusions before a research direction is closed.
EvoMap frames the release as a step toward AI4AI, extending later to drug discovery and materials science
EvoMap positions AutoResearch within a broader shift from AI assisting research to AI testing its own ideas against evidence, with potential extension beyond AI development into drug discovery, materials science and engineering domains where hypotheses can be evaluated through simulation or physical experiment. The accompanying paper, "AutoResearch: Insight In, Hallucination Out," is posted on arXiv, and the code is available on GitHub for researchers building AI scientists, autonomous agents and AI4AI systems. EvoMap's other projects include the Genome Evolution Protocol and EvoX Agent.