A paper introduces Lab2Sim, a benchmark for turning science experiment videos into interactive virtual experiments, plus a method called EduVideo2Code.
Benchmark: 218 science experiment videos, with 11,881 rubric items.
Method: EduVideo2Code extracts procedures hierarchically, encodes them as state machines to guide code generation, and refines the pages through execution-based verification.
Results: Existing multimodal models readily generate runnable pages but struggle to rebuild the apparatus faithfully. EduVideo2Code improves all four evaluation dimensions, with apparatus fidelity gains of 24.15 points.
Sign in to comment.
TL;DR of the paper, generated automatically.
A paper introduces Lab2Sim, a benchmark for turning science experiment videos into interactive virtual experiments, plus a method called EduVideo2Code.