LabEvolver: Training-Free Framework Enhances Wet-Lab Agents with Episodic Memory

A new framework called LabEvolver has been developed by researchers to provide wet-lab agents with episodic memory based on their execution experiences, all without the need for training. This system features an inner trial loop that is state-grounded, allowing for adaptive perception, online planning, and safety validation, paired with an outer evolution loop that transforms completed trajectories into reusable skills, strategies, and safety insights. In robotic solution-preparation tasks, LabEvolver proved its effectiveness by cutting down pH-regulation completion time by 48.2% and decreasing safety-gate intercepts by 60.0%. Additionally, it enhanced the cumulative success rate on the ALFWorld benchmark from 76.2% with ReAct to 91.4% across 500 continual tasks, demonstrating its applicability beyond wet-lab environments. More information can be found at https://andygao6186.github.io/LabEvolve.