LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents
TLDR
A training-free framework using episodic memory and experience evolution to improve wet-lab and household task agents.
Reasoning
The paper presents a novel training-free framework with strong empirical results on real-world robotic tasks and a simulated benchmark, demonstrating clear improvements. However, the abstract lacks detailed discussion of limitations, such as generalizability beyond the tested tasks or potential failure modes, and the core contribution is narrowly focused on experience evolution rather than broader automated scientific discovery.
Read-first score
Read-first score 51.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 47.
Field roles
Rank sensitivity
Stability: volatile; rank range: 43.
Keyword Scores
Deep Analysis
Innovations
- Training-free framework for equipping wet-lab agents with episodic memory from execution experience.
- State-grounded inner trial loop for adaptive perception, online planning, and safety validation.
- Outer evolution loop that distills completed trajectories into reusable skill, strategy, and safety experience.
- Demonstration of real-world feasibility on robotic solution-preparation tasks and generality on ALFWorld.
Methodology
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. It is evaluated on robotic solution-preparation tasks and the ALFWorld benchmark, comparing against ReAct.
Key Results
On robotic solution-preparation, LabEvolver reduced pH-regulation completion time by 48.2% and safety-gate intercepts by 60.0%. On ALFWorld, it improved cumulative success rate within 20 steps from 76.2% (ReAct) to 91.4% over 500 continual tasks.