OPERA uses operator–residual feedback to constrain autonomous optical experiments

OPERA adds physically interpretable operator and residual feedback to a language-model agent controlling optical experiments. Across 270 author-designed test problems, the authors report reaching and maintaining target performance in 75.8% of cases, versus 33.9–60.1% for reference strategies, while using less of the allotted experimental budget. All metrics are author-reported.

Impact: structured physical feedback may reduce reward hacking in autonomous laboratories. The work is a preprint and does not establish independent replication, generalisation to other apparatus or unattended safety.