Multimodal Data Demos

A detailed text description plus reference images, rebuilt as an editable Blender asset or scene: the best of several agent rollouts, with the full trajectory.

Natural-language requests for browser games and three.js scenes, each with full reasoning and one runnable single-file HTML.

Drawings or rendered views to CadQuery parts and assemblies, scored against the reference geometry, with four recorded model runs per case.

Terminal tasks built from production drawings: part, assembly and drawings-to-machine, graded in a separate verifier container.

Real-world photos and figures with questions that chain counting, reading and mirror reasoning into one exact number; human-written reasoning.

Spot-the-difference, mazes, cube nets, shadows, counting and pattern puzzles with short exact answers and reference reasoning.

Expert questions on real scientific, engineering, medical and financial charts that require exact readings and calculation.

Professional questions over long public PDFs, each with a reference answer, an itemized rubric and page-anchored evidence.