LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents

2026-07-30

AI Summary

This study addresses the challenge of endowing wet-lab agents with safe, reliable, and evolvable execution capabilities without requiring training. The authors propose a training-free framework comprising an inner loop that enables adaptive perception, online planning, and real-time safety verification through state-anchoring circuits, and an outer loop that distills execution trajectories into reusable skills, strategies, and safety knowledge via an experience evolution circuit. This approach establishes the first closed-loop experience evolution mechanism for wet-lab agents. In real-world solution preparation tasks, it reduces pH adjustment time by 48.2% and decreases safety violations by 60.0%. Furthermore, on 500 consecutive tasks in ALFWorld, the success rate within 20 steps improves from 76.2% to 91.4%.

Abstract

We introduce LabEvolver, a training-free framework that equips safe and grounded wet-lab agents with episodic memory from execution experience. LabEvolver couples a state-grounded inner trial loop for adaptive perception, online planning, and safety validation with an outer evolution loop that distills completed trajectories into reusable skill, strategy, and safety experience. On robotic solution-preparation tasks, LabEvolver demonstrates real-world feasibility, reducing pH-regulation completion time and safety-gate intercepts by 48.2% and 60.0%, respectively. On ALFWorld, it further improves cumulative success rate within 20 steps from 76.2% with ReAct to 91.4% over 500 continual tasks, showing generality beyond wet-lab settings. These results support learn-by-doing experience evolution as a feasible path toward closed-loop automated scientific discovery. The project page is available at https://github.com/AndyGao6186/LabEvolver.

Problem

• wet-lab agents
• training-free
• experience evolution
• safety validation
• automated scientific discovery

Innovation

• training-free
• experience evolution
• wet-lab agents
• safety validation
• episodic memory