SwarmRSI matches benchmark SOTA on 44.4% of NatureBench tasks

A Huawei-led team reports that its scientific agent swarm matched the published state of the art on 44.4% of the 90 tasks in NatureBench, the highest Match-SOTA rate among the methods it compared itself against. The preprint, posted to arXiv on October 14 with authors from Huawei Technologies, Leiden University, and the University of Groningen, calls the system SwarmRSI.

The pitch is that a research swarm should carry forward not just prompts but what one agent learned. The abstract describes a framework in which "a swarm improves its research by evolving the research states its agents inherit, including hypotheses, programs and experimental evidence." Three mechanisms do the work: Feedback-Guided Swarm Mutation decides whether to rewrite a hypothesis or its code after another agent's feedback; Complementary Swarm Recombination tests retained hypotheses on other agents' programs; Recursive Swarm Evolution feeds every outcome back into a shared graph.

With Gemini 3.1 Pro Preview, the paper reports an average score of 76.2 on OpenProblems. On MLE-Bench Lite, "it wins medals on 81.8% of tasks, above hypothesis and program evolution agents with the same backbone." In a single seed run with DeepSeek-V4-Pro across all 90 NatureBench tasks, "about 77% of final best programs descend from a rebase, a crossover or an update triggered by feedback." The authors offer that figure as evidence that the evolution mechanics, not just the backbone, are doing the work.

The out-of-benchmark claim is thin but concrete. "Biologists reproduce two of its programs and use them in follow-up analyses of intestinal and splenic immunity," the paper reports. It does not name those biologists or their labs.