Source record

Published by arXiv cs.RO (T1 source). The original is at https://arxiv.org/abs/2607.27690.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

LabEvolver is a training-free framework for wet-lab agents that pairs an inner trial loop for adaptive perception, online planning, and safety validation with an outer loop distilling completed trajectories into reusable skills, strategies, and safety experience. It cuts pH-regulation time and safety-gate intercepts on robotic solution-preparation tasks by 48.2% and 60%, and raises ALFWorld cumulative success within 20 steps from 76.2% (ReAct) to 91.4% over 500 continual tasks.

Training-free experience evolution with concrete safety and success gains across wet-lab and embodied benchmarks — a transferable pattern for memory-augmented agents.

Cited by

No citations on record.