Research map / Large language models / Agents, code and math
Targeted control of fast prototyping through domain-specific interface
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an LLM-based domain-specific interface (DSL) that translates designers' natural-language intentions into FreeCAD modeling commands for targeted control of fast prototyping, with automated domain adaptation via MCMC sampling and DPMM clustering.
Problem
Designers' natural-language instructions for adjusting prototype models are poorly translated into low-level CAD modeling commands due to gaps in abstraction level, semantic precision, and domain specificity.
Approach
An intermediate-level, domain-specific interface DSL that binds design concepts to permissible operations, automatically specified per product domain by LLM-driven MCMC sampling, DPMM clustering, and an EM-like reciprocal optimization validated against the FreeCAD API.
Key results
- Our-Int significantly outperforms LLM-Int in rendering consistency (t(7183)=80.60, p<.0001) and w/o-Int (t(7246)=63.81, p<.0001).
- Our-Int beats LLM-Int and w/o-Int in information clarity across eight domains (t(14)=4.413, p<.001; t(14)=3.277, p<.01).
- Domain adaptation costs about $10 per domain and executing the interface costs $0.3 per ten refinement iterations using GPT-4o.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the hierarchical DSL interface instead of a prompt-engineered LLM interface | LLM-based CAD generation · Eight product design domains | Rendering consistency: t(7183)=80.60, p<.0001 | Helped |
| Add the interface instead of processing designers’ language directly | LLM-based CAD generation · Eight product design domains | Rendering consistency: t(7246)=63.81, p<.0001 | Helped |
| Use a prompt-engineered LLM interface instead of processing designers’ language directly | LLM-based CAD generation · Eight product design domains | Rendering consistency: t(7127)=12.37, p<.0001 | Helped |
| Use the hierarchical DSL interface instead of a prompt-engineered LLM interface | LLM-based CAD generation · Eight product design domains | Information clarity: t(14)=4.413, p<.001 | Helped |
| Add the interface instead of processing designers’ language directly | LLM-based CAD generation · Eight product design domains | Information clarity: t(14)=3.277, p<.01 | Helped |
Surprising findings
- Increased complexity of generated modeling programs raised run-time error probability in the modeling engine, though the authors consider this minor given LLM-based program fixers.
Limitations
- Interaction modeled as a Markov process conditioned only on the last step.
- Only natural-language instructions are supported as input medium; sketches or gestures are not.
- Creating ground-truth modeling commands requires considerable expert effort.
Experimental setup
Datasets: LLM-generated design concept samples (teapot, pan, fork, toaster, cleaver knife, sofa, wheelbarrow, unicycle)
Benchmarks: human study rendering consistency rankings (50 participants, 8 product domains) information clarity via AST cumulative depth
Metrics: rendering consistency (ranked order) information clarity (AST depth) paired t-test
Baselines: w/o-Int LLM-Int
Builds on: Metropolis-Hastings Dirichlet process mixture model FreeCAD
Closest papers by meaning
- FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models (ICLR 2025 · similarity 0.83)
- CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.82)
- CAD-Editor: A Locate-then-Infill Framework with Automated Training Data Synthesis for Text-Based CAD Editing (ICML 2025 · similarity 0.82)
- Toward Engineering AGI: Benchmarking the Engineering Design Capabilities of LLMs (NeurIPS 2025 · similarity 0.82)
- CANVAS: A Benchmark for Vision-Language Models on Tool-Based User Interface Design (AAAI 2026 · similarity 0.81)
- Position: You Can't Manufacture a NeRF (ICML 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.