SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm

Wed, 14 Oct 2026 00:00:00 +0000

Motivation

Sharing experience across a scientific agent swarm does not by itself lead to better research decisions. A poor result does not show whether the hypothesis or its implementation needs revision, so promising hypotheses get abandoned; and a program without overall gains may still hold components useful to another agent’s later research.

Two limitations motivating SwarmRSI.

Method

SwarmRSI evolves the research states that agents inherit — hypotheses, programs, experimental evidence and unresolved needs — through three mechanisms:

Feedback-Guided Swarm Mutation uses other agents’ feedback to decide whether to revise a hypothesis or repair its program.
Complementary Swarm Recombination tests retained hypotheses on other agents’ programs (rebase) and combines complementary program components (crossover).
Recursive Swarm Evolution returns every outcome to a shared research graph that guides which state each agent updates next.

BibTeX

@misc{meng2026swarmrsirecursiveselfimprovementevolving,
 title={SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm},
 author={Qianru Meng and Xi Zhang and Konstantinos Papakostas and Congfeng Cao and Wenxiang Hu and Tianyi Lu and Xiao Zhang and Lun Zhou and Zaiqiao Meng and Zhaochun Ren and Xiaohui Yan and Yougang Lyu},
 year={2026},
 eprint={2610.20431},
 archivePrefix={arXiv},
 primaryClass={cs.AI},
 url={https://arxiv.org/abs/2610.20431},
}

Recursive Self-Improvement | Xi Zhang

SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm

Motivation

Method

BibTeX