Research map / Vision-language and multimodal / Human and embodied modeling
Text-to-Code Generation for Modular Building Layouts in Building Information Modeling
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.
Text2MBL fine-tunes Qwen2.5 LLMs to translate textual descriptions of modular building layouts into executable Revit BIM code, outperforming coordinate-based generation on a 198-design dataset augmented with synthetic data.
Problem
Generating parametric, semantically rich BIM layouts for modular building layouts (modules, units, rooms) directly from natural language, where existing methods output 2D or bounding-box coordinates that are error-prone and lack executability.
Approach
Reframe text-to-BIM layout generation as text-to-code: an object-oriented Revit C# API (Module, Unit, Room, Utils classes) is exposed to LLMs, which output action sequences of relative-position constructor calls that execute directly in Autodesk Revit.
Key results
- Qwen2.5-Coder-7B code output reached 98.43% IoU versus 84.64% for coordinate output with the same model size.
- Qwen2.5-Coder-7B with positional arguments achieved 99.0% compile rate and 79.25% pass rate.
- Mixing synthetic data gave substantial gains in low-resource settings (e.g., only 10% of original data available).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Generate BIM code instead of coordinate sequences | Qwen2.5-Instruct 7B · MBL test | IoU (%): 85.38 -> 95.83 | Helped |
| Generate BIM code instead of coordinate sequences | Qwen2.5-Coder-Instruct 7B · MBL test | IoU (%): 84.64 -> 98.43 | Helped |
| Generate BIM code instead of coordinate sequences | Qwen2.5-Math-Instruct 7B · MBL test | IoU (%): 85.34 -> 94.86 | Helped |
| Increase model scale from 1.5B to 7B | Qwen2.5 model families · MBL test | IoU (%): 1.5B - Code: 91.65 -> 95.83; 94.19 -> 98.43; 94.18 -> 94.86 | Helped |
| Use positional rather than named arguments | Qwen2.5 model families · MBL test | Compile (%) | Helped |
| Switch between named and positional arguments | Qwen2.5 model families · MBL test | Argument F1 (%) | No clear effect |
| Augment original training data with synthetic data | Qwen2.5-Coder-3B · MBL test, especially 10% original data | Performance across evaluation metrics | Helped |
| Train on synthetic data alone rather than augmenting original data (ablation) | Text2MBL · MBL test | Performance | Hurt |
| Pretrain on synthetic data before fine-tuning on original data (ablation) | Text2MBL · MBL test | Performance | No clear effect |
| Use named rather than positional arguments | Qwen2.5-Coder-3B · Abstract instructions | Compile (%): 45.00 -> 68.75 | Helped |
| Use named rather than positional arguments | Qwen2.5-Coder-7B · Abstract instructions | Compile (%): 95.00 -> 92.5 | Hurt |
Surprising findings
- Argument F1 and instance F1 were notably higher than pass rate, so models extract information better than they produce fully executable code.
- Proprietary gpt-4.1 generated coordinates with only 29.94% IoU, and gpt-4.1-mini code had a pass rate of 6.25%.
Limitations
- Only four room types (living room, bedroom, bathroom, kitchen) and rectangular modules are supported.
- Pass rate remains low for modules and rooms; generated code often has wrong argument order or hallucinated functions.
- Small dataset of 198 designs, and user-centric evaluation (satisfaction, usability) was not performed.
Experimental setup
Datasets: Hong Kong pre-accepted modular integrated construction designs (198 MBL designs) synthetic descriptions (template- and model-based) fully synthetic code (3,000 sequences)
Benchmarks: Text2MBL test set (40 designs, 80 pairs)
Metrics: compile rate pass rate instance F1 argument F1 IoU
Baselines: coordinate-driven generation (Qwen2.5 bounding box output) ChatGPT gpt-4.1 ChatGPT gpt-4.1-mini
Builds on: LoRA AutomaTikZ Tell2Design
Closest papers by meaning
- CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation (CVPR 2025 · similarity 0.85)
- LayoutNUWA: Revealing the Hidden Layout Expertise of Large Language Models (ICLR 2024 · similarity 0.84)
- Tokenization Allows Multimodal Large Language Models to Understand, Generate and Edit Architectural Floor Plans (CVPR 2026 · similarity 0.83)
- CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules (ICLR 2024 · similarity 0.82)
- Text2City: One-Stage Text-Driven Urban Layout Regeneration (AAAI 2024 · similarity 0.82)
- FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models (ICLR 2025 · similarity 0.82)
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.