Description

We study AI-agent methods that reconstruct, generate, or refine 3D structures and time-varying 3D models for scientific inference, physical simulation, or function-driven design, evaluated beyond visual appearance.

Scientific simulation and design require 3D/4D models whose geometry, motion, and physical properties support reliable predictions. The third 3D4S workshop explores AI agents that reconstruct, generate, and refine these models through measurement, geometric verification, and numerical simulation. We bring together computer vision, graphics, scientific computing, and domain science to study tool orchestration, uncertainty, physical validity, and evaluation beyond visual appearance. Through invited talks, contributed papers, posters, and discussions, participants will develop shared evaluation criteria and reproducible workflows for agentic modeling in engineering, environmental science, biomedicine, and materials research.

What do we mean by agentic? An agent selects or revises modeling actions using observations, tool outputs, or feedback. We welcome tool-using language and vision-language agents, learned controllers, and human-agent collaboration. 4D means time-varying 3D, including geometry, motion, and physical properties.

The scientific question comes first: a force, flux, deformation, material property, or functional constraint. This raises three questions:

• Q1: How can agents integrate heterogeneous images, depth, tomography, and trajectories to reconstruct reliable 3D/4D models and identify missing evidence?
• Q2: How can scientific feedback guide geometry generation, meshing, physical identification, and inverse design toward measurable physical objectives?
• Q3: How can we evaluate accuracy, uncertainty, and computational efficiency, including the benefit of adaptive agents over fixed workflows with matched resources?

Building on WACV 2024 and CVPR 2026, our third edition extends scientific geometry reconstruction and generation to agent-directed modeling: observe or specify, build, verify, simulate or measure, diagnose, and revise. CVPR 2026 brought together 150 attendees and received 21 paper submissions.