SwarmRSI：透過演進式科學智能體羣實現遞歸自我改進
SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm
SwarmRSI 把成功、失敗及無增益結果回饋共享研究狀態圖，藉此更新科學智能體羣後續工作的起點。在118項科學編程及 ML 工程任務中，OpenProblems 平均分為76.2，高於人類專家（57.2）及 TusoAI（67.1）。NatureBench 90 項任務中，44.4% 達已發表 SOTA，為比較方法中最高。
Something I keep noticing with AI scientist agents: running more of them in parallel doesn’t make them smarter by default. Mostly they throw away each other’s work. 🧪
Picture a lab where a student has a promising idea but botches the implementation, and the idea gets crossed off. Down the hall, another student’s “failed” project holds the one trick a colleague needed, and nobody ever opens it.
Agent swarms run into the same two walls:
• A failed experiment doesn’t tell you whether the hypothesis was wrong or the code was
• A program with no overall gain can still hold a component another agent needs
• So good ideas get abandoned and useful pieces get buried
That’s the problem behind our new paper:
🚀 SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm
The idea: let the swarm evolve what its agents inherit, not just keep its best scores.
• A shared graph records each research state: hypotheses, programs, experimental evidence and open needs
• Feedback-Guided Swarm Mutation uses other agents’ feedback to decide whether to revise the hypothesis or repair the program
• Complementary Swarm Recombination re-tests kept hypotheses on other agents’ programs (rebase) and merges components that complement each other (crossover)
• Recursive Swarm Evolution sends every outcome back into the graph, whether it succeeded, failed or made no gain, so it shapes where the next agent starts
Results on 118 scientific coding and ML engineering tasks:
📈 76.2 average on OpenProblems (single-cell), vs 57.2 for human experts and 67.1 for TusoAI
🧬 Matches published SOTA on 44.4% of NatureBench’s 90 tasks, the highest of all compared methods (the best Claude Code setup reaches 35.6%)
🏅 Medals on 81.8% of MLE-Bench Lite competitions, above Arbor and MLEvolve
🔁 About 77% of final best NatureBench programs come from a rebase, a crossover or a feedback-triggered update
🔬 Biologists reproduced two of its programs and used them in follow-up studies of intestinal and splenic immunity
I think the next progress in AI for science will come less from adding agents or scaling models, and more from making sure what one agent learns changes where the next one begins.
📝 Paper: https://arxiv.org/abs/2610.20431
來源：HuggingFace Daily Papers（社區熱門論文） · huggingface.co