Skip to content
Bud's Wiki

SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm

TL;DR

Section titled “TL;DR”

SwarmRSI frames recursive self-improvement (RSI) in scientific research as an evolving swarm of agents rather than a single autoresearcher: each agent carries a hypothesis + implementation pair, and the swarm improves through three evolution operators — Feedback-Guided Swarm Mutation (refine hypotheses and code using feedback from other agents), Complementary Swarm Recombination (reuse retained hypotheses and program components across agents so partial results support new research), and Recursive Swarm Evolution (feedback loop binding both into a continuing cycle). The authors report SOTA on OpenProblems, NatureBench, and MLE-Bench Lite, with accumulated research experience transferring to held-out tasks, and claim biologists have reproduced its programs for follow-up analyses.

Key claims

Section titled “Key claims”
  • Scientific agent swarms benefit from being able to combine evidence and implementations from different research directions to guide subsequent experiments, but naive overall program scores hide useful components and poor results leave ambiguity between hypothesis vs implementation failure [§Abstract].
  • Feedback-Guided Swarm Mutation refines hypotheses and implementations using experimental feedback from other agents (not just the agent’s own failures), addressing the hypothesis-vs-implementation ambiguity by sharing cross-agent diagnostic signal [§Abstract].
  • Complementary Swarm Recombination reuses retained hypotheses and complementary program components across agents, so partial results — even from runs that didn’t win overall — can support new research directions [§Abstract].
  • Recursive Swarm Evolution composes mutation + recombination into a feedback loop where new experimental evidence guides subsequent research rounds, framing swarm evolution as the RSI substrate rather than a single agent’s rewrite loop [§Abstract].
  • Reports SOTA performance on OpenProblems, NatureBench, and MLE-Bench Lite [§Abstract].
  • Accumulated research experience improves performance on held-out tasks — a transfer claim, not just an in-distribution win [§Abstract].
  • Biologists reproduce SwarmRSI’s programs for follow-up biological analyses, implying the generated code clears a human-expert replication bar in at least one domain [§Abstract].

Method

Section titled “Method”

SwarmRSI maintains a swarm of scientific-research agents, each holding a (hypothesis, implementation) pair. Three evolution operators cycle: (1) Feedback-Guided Swarm Mutation uses experimental feedback from other agents in the swarm to refine both the hypothesis and its code implementation, giving the mutation operator cross-agent diagnostic signal rather than only within-agent trial-and-error; (2) Complementary Swarm Recombination identifies retained hypotheses and complementary program components that can be stitched together across agents, so useful fragments from losing runs still propagate; (3) Recursive Swarm Evolution wraps mutation + recombination into a feedback loop, using new experimental evidence as input to the next round. Evaluation is on OpenProblems, NatureBench, and MLE-Bench Lite, with a held-out task set used to measure transfer of accumulated research experience.

Results

Section titled “Results”
  • SOTA on OpenProblems, NatureBench, and MLE-Bench Lite [§Abstract].
  • Accumulated research experience improves performance on held-out tasks (transfer, not just in-distribution) [§Abstract].
  • Biologists reproduce SwarmRSI’s generated programs for follow-up biological analyses [§Abstract].

(Specific numbers, baselines, and ablations are not visible from the abstract; the paper text is the authoritative source.)

Why it’s interesting

Section titled “Why it’s interesting”

Swarm-based RSI occupies a distinct cell on the AI-for-AI-research granularity axis: unlike AIDE²: First Evidence of Recursive Self-Improvement (AIDE² rewrites a single inner agent) or Dream-RSI: Recursive Self-Improvement through Evolving Worlds (Dream-RSI meta-optimizes a single exploration policy via off-policy replay), SwarmRSI’s unit of evolution is a population of (hypothesis, implementation) pairs with cross-agent mutation and recombination — closer in spirit to evolutionary program synthesis (AlphaEvolve-family) than to the single-trajectory autoresearch scaffolds that dominate AI-for-AI Research filings. The Complementary Swarm Recombination operator directly addresses the “losing runs still hold useful fragments” problem that Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents (Fork-and-Flush) tackles via workspace-sharing + context resets — both are counter-recipes to best-of-N that preserve partial signal, but SwarmRSI does it structurally (component-level reuse) rather than schedule-wise (fork intervals).

See also

Section titled “See also”