pipette
ESEspañol

SEA-Lab: Experiment-grounded recursive self-improvement in an automated laboratory

Y. Jiang, R. Gong, H. Chen, X. Wang, B. Zhang, Q. Wang, T. Zhang, M. Fan, X. Mao, Z. Hou, Y. Hu, Y. Gao, C.-Z. Wang

Preprint

In the authors' words

AI scientists can generate hypotheses, plans, and protocols. Translating these outputs into reliable experiments requires adapting them to local equipment and resources using feedback from observed outcomes. A key challenge is to retain laboratory execution feedback as reusable knowledge for subsequent planning. We present SEA-Lab (Self-Evolving Autonomous Lab), which combines an environment-grounded Dry-Lab Harness that generates protocols and machine-readable procedures, a wet-lab execution harness that coordinates heterogeneous devices and records execution traces and experimental measurements, and a recursive self-improvement mechanism that uses physical feedback to revise reusable execution policies and experimental strategies. Across three rounds of review on 219 tasks, protocol-review pass rates reached 90.41%. On 66 automation-eligible tasks, mean virtual-validation pass rates increased from 16.7% to 87.7% after two revisions. On two liquid handlers, runtime fell by 45.8% for plate replication, 34.1% for well rearrangement, and up to 68.6% for sample normalization. We demonstrated the system in a phage-display experiment linking four automated experimental steps: molecular cloning, phage titration, sample normalization, and ELISA screening. Of 96 starting samples, 93 yielded a sequence-correct clone, and primary ELISA screening identified ten candidates. These results illustrate how laboratory feedback can guide planning around local resources, refine execution on different laboratory instruments, and inform method revision based on experimental results.

Main resultThe abstract does not state a limitation.

In plain words AI summary

The authors built SEA-Lab, a system that lets AI scientists plan experiments, run them on real lab equipment, and learn from the results. In a phage-display experiment with four automated steps, 93 of 96 samples produced correct clones. The results show how lab feedback can guide planning, refine execution on different instruments, and improve methods based on outcomes.

Written by Kimi K2.5 using only the abstract above. Jev checked that every sentence is supported by the abstract sentences it cites, and removed any that were not. Point at a sentence to highlight its source. 2 sentences were removed for not being supported by the abstract.

Appeared: Tuesday, November 17. bioRxiv. Preprint, not yet peer-reviewed.

DOI: 10.64898/2026.11.12.766090