SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm

SwarmRSI. Recursive Self-Improvement through Evolving Scientific Agent Swarm. By Qianru Meng and eleven others, from Huawei, Leiden University, and one other institution. Published on the fourteenth of October. No code repository is linked to it.

Recursive self-improvement in scientific research means agents getting better at research through their own experimental experience. A swarm of scientific agents can help, by pooling evidence and implementations from different research directions to guide what each agent tries next. Two things get in the way. When an experiment goes badly, it can be unclear whether the hypothesis was wrong or only its implementation. And a single overall score for a program can hide components inside it that other agents would have found useful.

SwarmRSI is a framework that evolves scientific agent swarms for recursive self-improvement, and it answers each of those two problems with its own mechanism. Then it joins the two into a loop.

Feedback-Guided Swarm Mutation addresses the first problem. It refines both hypotheses and their implementations, using experimental feedback that comes from other agents. Complementary Swarm Recombination addresses the second. It reuses hypotheses that were retained, along with complementary program components, across agents, so that partial results can support new research instead of being discarded with a weak overall score. Recursive Swarm Evolution integrates both mechanisms into a feedback loop, in which new experimental evidence guides the research that follows.

Experiments on OpenProblems, NatureBench and MLE-Bench Lite show state-of-the-art performance for SwarmRSI. Further evaluations show that the research experience it accumulates improves performance on held-out tasks. And biologists have reproduced its programs for follow-up biological analyses.

Each mechanism maps onto one of the two obstacles. A poor result no longer leaves the agent guessing in the dark, because mutation can refine the hypothesis, the implementation, or both, informed by what other agents have observed. A program with an unremarkable overall score no longer takes its good components down with it, because recombination can lift those components into another agent's work. The swarm keeps the pieces, the loop carries them into the next round, and the held-out results suggest the experience gathered this way travels beyond the tasks it came from.