Agentic Creation & Executable Artifacts
Agents that transform instructions into runnable games, CAD scripts, mathematical solvers, optimization models, and executable reasoning traces.
Agentic Creation & Executable Artifacts asks whether models can create things that actually run. A strong agent should not only produce plausible text; it should deliver a complete project, script, model, solver, trace, or environment that can be launched, executed, replayed, rendered, and verified.
Research Storyline
GameCraft-Bench evaluates full Godot projects through launchability, replay traces, gameplay evidence, and hidden rubric judging.
BlenderLLM generates Python CAD scripts that execute inside Blender and render 3D models from user instructions.
Mamo and ORLM connect natural language problems to formal models, constraints, objective functions, solvers, and benchmarkable outputs.
CoRT integrates code execution into the thinking process so intermediate calculations can be checked rather than only narrated.
Representative Work
Tests whether agents can build complete playable games in a real game engine with replay-based verification.
PaperTrains LLMs for CAD script generation with BlendNet, CADBench, public models, and self-improvement.
PaperBenchmarks mathematical modeling with solver-backed executable answers.
PaperTrains large models for automated optimization modeling and releases benchmark resources for industrial OR problems.
PaperConnects reasoning to executable computation, making intermediate reasoning more testable.
PaperWhy It Matters
- It shifts evaluation from static answers to complete artifacts that can be launched, rendered, replayed, or solved.
- It gives agent research stronger evidence: traces, states, runtime behavior, and verifier outputs.
- It connects coding agents, creative generation, CAD, optimization, and reasoning under one executable-artifact agenda.