Recursive self-improvement (RSI) is a long-standing goal in AI: a system that gets better at research by doing research. The new arXiv preprint SwarmRSI applies that idea to scientific agent swarms, where multiple agents work together and build on each other's findings. According to the abstract, the approach aims to improve agents' research capabilities through their own experimental experience.

The key mechanism is combining evidence and implementations from different parts of the swarm. That suggests a collective process where one agent's results and code can inform another's next steps, allowing the swarm to leverage distributed discoveries. The framing is explicitly scientific, pointing toward tasks like experimental design or hypothesis testing.

The abstract is brief and does not yet report quantitative results, so it is hard to judge how well SwarmRSI works in practice. Still, it points to a promising direction: using the swarm itself as the training signal for recursive improvement.