Separate papers advance scientific swarms and ML agents

SwarmRSI claims SOTA on research benchmarks, while 35B MLExpert lifts MLE-Bench Lite medal rate from 37.9% to 72.7%.

Two separate papers, not one system, describe agentic approaches to scientific research and machine-learning engineering. SwarmRSI uses feedback-guided mutation of hypotheses and implementations, recombination of retained hypotheses and program components, and recursive evolution on new evidence; its authors report state-of-the-art results on OpenProblems, NatureBench, and MLE-Bench Lite, with further gains on held-out tasks from accumulated experience. MLExpert is a 35B mixture-of-experts model trained with hierarchical agentic supervised fine-tuning and Windowed Interaction RL on long experiment trajectories built with Kaggle-MLE Factory. On MLE-Bench Lite it lifts the base model's average medal rate from 37.9% to 72.7%, a level the reports call comparable to larger models such as Kimi-K3 and GLM-5.3, with additional gains on NatureBench Lite, agent-scaffold optimization, and algorithm optimization. The two MLExpert writeups agree on model size and medal rates; the later arXiv summary adds that asynchronous collection allows reward updates before a trajectory finishes. Sources do not equate SwarmRSI's NatureBench result with MLExpert's NatureBench Lite gains.

• SwarmRSI evolves scientific agent swarms via Feedback-Guided Swarm Mutation, Complementary Swarm Recombination, and Recursive Swarm Evolution.
• Authors report SwarmRSI state-of-the-art on OpenProblems, NatureBench, and MLE-Bench Lite, plus held-out gains from accumulated experience.
• MLExpert is a 35B mixture-of-experts agent for long-horizon autonomous machine-learning engineering.
• Training uses hierarchical agentic supervised fine-tuning and Windowed Interaction RL; Kaggle-MLE Factory supplies executable tasks.
• On MLE-Bench Lite, MLExpert raises the base model's average medal rate from 37.9% to 72.7%, comparable to larger models such as Kimi-K3 and GLM-5.3.
• MLExpert also improves NatureBench Lite, agent-scaffold optimization, and algorithm optimization.
• The Oct. 15 arXiv note adds that asynchronous collection lets Windowed Interaction RL update from outcomes before a trajectory ends.