Le Chen, Nuo Xu, Winson Chen, Bin Lei, Pei-Hung Lin, Dunzhi Zhou, Rajeev Thakur, Caiwen Ding, Ali Jannesari, Chunhua Liao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Photonic waveguide simulation in Ansys Lumerical Finite-Difference Time-Domain (FDTD) requires translating high-level design intent into executable simulator programs and physically meaningful experimental setups. This translation is difficult for general-purpose language models because the Lumerical Python API is proprietary and setup quality depends on photonic judgment about geometry, materials, sources, monitors, mesh, and output interpretation. We present Lumi Agent, a closed-loop agentic workflow for generating and refining Lumerical waveguide simulation from natural-language requests. The workflow combines domain guidance, constrained file and solver actions, simulator feedback, and failure reflection so each revision is conditioned on what the previous run actually produced. To separate execution from scientific quality, we evaluate Lumi Agent on a 14-task waveguide benchmark using execution-log metrics and human-expert review of setup correctness, result correctness, and reasoning over underspecified task details. Direct first-attempt generation by GPT-5.2, Claude Sonnet 4.6, and Gemini 3.1 Pro produced 0/14 executable scripts for each model, whereas Lumi Agent achieved 14/14 executable scripts and clean simulation runs for each backend. Human review further identifies physics-level setup and result errors beyond clean execution. These results show that grounded simulator feedback makes Lumerical simulation generation more executable, inspectable, and repairable while keeping expert photonics review in the loop.

This paper introduces \textsc{InfantAgent-Next}, a generalist agent capable of interacting with computers in a multimodal manner, encompassing text, images, audio, and video. Unlike existing approaches that either build intricate workflows around a single large model or only provide workflow modularity, our agent integrates tool-based and pure vision agents within a highly modular architecture, enabling different models to collaboratively solve decoupled tasks in a step-by-step manner. Our generality is demonstrated by our ability to evaluate not only pure vision-based real-world benchmarks (i.e., OSWorld), but also more general or tool-intensive benchmarks (e.g., GAIA and SWE-Bench). Specifically, we achieve a $\mathbf{7.27\\%}$ accuracy gain over Claude-Computer-Use on OSWorld. Codes and evaluation scripts are included in the supplementary material and will be released as open-source.