Research map / Multi-agent systems and games / LLM agents

LensDesigner: A Self-Improving Agent for Optical Lens Design

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LensDesigner, an LLM agent framework with retrieval from a 100K-lens library, optical tools, macro orchestration and skill-library self-evolution for spherical imaging lens design, evaluated on a new 120-task LensArena benchmark.

Problem

Automated optical lens design is a non-convex, experience-dependent optimization problem; existing optimization, deep-learning and retrieval methods are slow, require manual tuning, or cannot handle novel specifications.

Approach

An LLM agent mirrors expert opticians: it starts from physically valid seeds retrieved by an optics-aware distance metric, edits lenses via guarded optical tools and Zemax DLS optimization, and accumulates design heuristics in a skill library through curriculum-driven self-evolution without weight updates.

Key results

• LensDesigner w/ Gemini-3.0-flash reaches 87.5% success on Easy and 80.0% on Hard LensArena tasks, versus 30.0% for the best prior method (LensNet) on Hard.
• On Hard tasks, LensDesigner w/ Gemini-3.0-flash attains Avg RMS 12.7 μm, compared with 7.99 μm for LensNet and 11.1 μm for QGSO (LensNet and QGSO averages are over successful designs only).
• Ablation with Gemini-2.5: replacing optics-aware retrieval with vanilla RAG raises Hard Avg RMS from 48.8 to 189.6 μm.

Surprising findings

• Vanilla in-context learning with frontier LLMs performs poorly on lens design, with Avg RMS errors often exceeding 100 μm.
• The Hard-split Avg RMS of LensNet (7.99 μm) is lower than LensDesigner's (12.7 μm) with Gemini-3.0-flash, though LensNet succeeds far less often.

Limitations

• Limited to spherical lenses; aspheric and freeform surfaces not supported.
• Latency from reliance on external Zemax ray tracing.
• Avg RMS computed only over successful designs, which can make comparisons across methods with different success rates misleading.

Experimental setup

Datasets: LensLib100K

Benchmarks: LensArena

Metrics: success rate average RMS spot radius (μm)

Baselines: LensNet QGSO DeepLens ICL with Gemini-3.0-flash ICL with ChatGPT-5.2 ICL with Kimi-2.4 ICL with Qwen-3-8b ICL with Qwen-3-32b

Builds on: QGSO ReAct Damped Least Squares

Closest papers by meaning

• Simple Agents Outperform Experts in Biomedical Imaging Workflow Optimization (CVPR 2026 · similarity 0.82)
• Frontier-Eng: Benchmarking Self-Evolving Agents on Real-World Engineering with Generative Optimization (NeurIPS 2026 · similarity 0.82)
• DesignX: Human-Competitive Algorithm Designer for Black-Box Optimization (NeurIPS 2025 · similarity 0.81)
• Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.81)
• AutoMS: Multi-Agent Evolutionary Search for Cross-Physics Inverse Microstructure Design (ICML 2026 · similarity 0.80)
• PhotoFlow: Agentic 3D Virtual Photography Missions (NeurIPS 2026 · similarity 0.79)