Research paper 2026
LensDesignerA Self-Improving Agent for
Optical Lens Design
Learning the craft of optical design, one experience at a time.
1 INSAIT, Sofia University “St. Kliment Ohridski”2 Zhejiang University
The idea
Optical design that learns from experience.
Optical lens design relies on expert intuition to explore a complex space of shapes, materials, and physical constraints. LensDesigner brings this experience-driven process to an autonomous agent.
The framework retrieves physically valid starting structures from LensLib100K, refines them through feedback from an optical simulator, and stores reusable design heuristics. A curriculum agent proposes increasingly difficult tasks so the system can build its expertise over time. We evaluate it on LensArena, a benchmark of 120 optical design tasks spanning common and extreme configurations.
Retrieve a starting point
Optics-aware retrieval selects structural seeds from LensLib100K.
Refine with physics
The agent orchestrates structural changes and local optimization with simulator feedback.
Build reusable skills
A curriculum guides experience accumulation across progressively harder tasks.
Explore the framework

If the preview is unavailable on your device, open the PDF directly.
If LensDesigner informs your research, please use the following citation.
@misc{sun2026lensdesigner,
title = {{LensDesigner}: A Self-Improving Agent for Optical Lens Design},
author = {Sun, Lei and Liang, Haoran and Xu, Dannong and Gao, Yao
and Geng, Yuyu and Gu, Jinjin and Wang, Kaiwei
and Paudel, Danda Pani and Van Gool, Luc},
year = {2026},
url = {https://Danielxu0208.github.io/LensDesigner/}
}