Yongqing Jiang1† · Haoran Luo1* · Jianze Wang2 · Xin Zhou1 · Kaoshan Dai2 · Zhiqi Shen1*
1Nanyang Technological University 2Sichuan University
†project leader *corresponding authors
[arXiv]
[Project page]
[Paper]
[BibTeX]
This work introduces the Vibe Building task and PE-Loop, an agent in which a deterministic physics engine, not a language model, decides whether a building design is accepted. The language model proposes discrete revisions and pays nothing for them; the physics engine holds the verdict and every decision costs one evaluation of the budget.
Note
This repository is the landing page for the paper. The code and VB-Bench are not public yet. See Release plan for what is coming and in what order. Watch or star the repository to be notified.
- 2026-10-06 Preprint on arXiv: arXiv:2610.08156
- 2026-10 Project page is live, with twenty-six buildings carried end to end through the pipeline: jovanqing.github.io/Vibe-Building
- 2026-10 Paper submitted. The PDF is available from the project page.
Most agents that design buildings produce models that look plausible but are never checked against mechanics or against code. We introduce the Vibe Building task and PE-Loop (Physics-Engine-in-the-Loop), an agent in which a deterministic physics engine is the only source of evaluation signal. The language model proposes discrete revisions and pays nothing for them. The physics engine decides, and every decision costs one evaluation of the budget. Replacing that verdict with a language-model judge, with everything else held fixed, leaves 58.43% of accepted designs noncompliant.
Automated building design must comply with seismic and wind codes and satisfy structural mechanics constraints, yet most existing agents produce visually plausible models without verification grounded in mechanical analysis and code compliance. We introduce the Vibe Building task and propose PE-Loop (Physics-Engine-in-the-Loop), an agent in which a deterministic physics engine is the sole source of evaluation signals, mapping code constraints to a physics process reward, while the language model is confined to proposing discrete revisions (section menu, topology, and lateral system). Designs are verified by held-out seismic and wind time-history checks and a constructability gate. On VB-Bench, 3,577 physics-adjudicated building instances across six code families, PE-Loop achieves the highest verified success rate under three of four backbone LLMs, the highest held-out seismic pass rate under all four, and the highest held-out wind pass rate under three. Replacing the physics verdict with a language-model judge, all else fixed, leaves 58.43% of accepted designs noncompliant. These results suggest that reliable structural design rests less on a stronger LLM proposer than on an adjudicator the proposer cannot influence, a division of labor for agents whose outputs must hold up in the physical world.
The meta level evolves discrete decisions (intent, system, strategy). The base level runs continuous optimisation and deterministic assembly. Physics verification returns parameter-level and topology-level feedback. The loop terminates only when every strength and serviceability constraint holds.
Every model below was produced by the pipeline and rendered by the same Blender scene, from four structural typologies that do not appear on the project page. The project page carries the other twenty-two, each with its Rhino massing, its OpenSeesPy frame and its first mode shape.
 Brutalist Castle Keep Tower · 106 m top-heavy overhang on a slender shaft |
 Diamond Slab Tower · 127 m slab tapering to a point at both ends |
 Brutalist Mega-Slab Residential · 59.5 m long slab lifted on pilotis |
 Art Deco Stepped Tower · 84 m setbacks narrowing toward the crown |
Five agent systems run under four backbone LLMs on the topology-decision instances of the main
split, at effort tier high. Every arm receives the same instances, the same margin oracle and
the same budget, so the only thing that varies is the system. Within each block the best value
per column is shaded blue and the second orange, and tied values share a mark. Lower is better
for NFE@p, Exh., Wall, T/O and Rank; ESC is reported but not ranked.
Column glossary (the paper gives full definitions and denominators in Appendix E)
| Column | Meaning |
|---|---|
| PASS | In-loop compliance: the design satisfies every strength and serviceability constraint. |
| ESC | Escaped: answers that left the loop without a verdict. Reported, not ranked. |
| NFE@p | Number of function evaluations spent to reach a pass. Lower is cheaper. |
| Exh. | Budget exhausted without a pass. |
| Wall | Wall-clock minutes per instance. |
| T/O | Timed out. |
| pass³ | Passes on all three seeds of the same instance. Stability under resampling. |
| Valid | Answer is a well-formed, parseable design. |
| Seis. / Wind | Held-out time-history analysis: seismic and wind records never seen in the loop. |
| Build | Constructability gate. |
| Verif. | Certified success: passes every gate above, including both held-out checks. |
| P/h | Passes per hour. Throughput. |
| Rank | Mean rank across the ranked columns within the block. |
± is the Agresti–Coull standard error (standard error of the mean for NFE@p). Wall is in minutes, P/h in passes per hour, every other column in percent. PASS, Seis., Wind, Build and Verif. are over the 390 topology-decision answers per arm; ESC, Exh., T/O and Valid over all 720 answers; pass³ over the 130 instances.
The verdict is handed to a language-model judge that sees the brief, the code limits and the design parameters but no simulator output. Everything else is fixed (Table 5 of the paper).
| Acceptance gate | Self-rep. PASS % | Verified PASS % ↑ | False-pos. % ↓ | Wasted NFE % ↓ | Jaccard ↑ | Wall s |
|---|---|---|---|---|---|---|
| Margin oracle (ours) | 77.78 | 77.78 | 0.00 | 0.00 | 1.00 | 110.8 |
| LLM judge (same LLM) | 49.44 | 20.56 | 58.43 | 44.41 | 0.31 | 238.4 |
| LLM judge (stronger, disjoint) | 40.00 | 17.22 | 56.94 | 45.19 | 0.29 | 342.8 |
The two zeros on the first row are zero by construction, not for want of data: the gate of the margin oracle is the physics check itself, so what it reports and what verification confirms are the same quantity.
Because the verdict never depends on the proposer, the held-out pass rate barely moves when the language model behind it changes. Spread across DeepSeek-V4-Flash, Qwen3.8-Flash, GLM 5.2 Fast and Kimi K3:
| PE-Loop | Best baseline | |
|---|---|---|
| Held-out seismic | 3.3 pp | 23.8 pp |
| Held-out wind | 4.1 pp | 27.7 pp |
3,577 physics-adjudicated building instances, 1,311 gradable tasks, six code families (GB, EC8, ASCE 7, BSL, IS, NZS). Every code limit is traceable to the clause it comes from.
Ordered against the ML Code Completeness Checklist. Nothing below is public yet; this section is the roadmap, and it is updated as items land.
- VB-Bench instances and the physics adjudication harness
- PE-Loop agent code (meta level, base level, margin oracle)
- Structural-code knowledge base: 86 codes, 625 cross-linked entries with clause provenance
- Evaluation code and the scripts that reproduce every table and figure in the paper
- Held-out record libraries: 44 FEMA P-695 far-field motions, 16 wind series
- Pipeline exporters: Rhino, Blender and OpenSeesPy artefacts shown on the project page
Release is gated on the review outcome. Opening an issue to ask about timing is welcome.
This page, the figures above and four turntable animations. The other twenty-two buildings, the full four-stage pipeline for each, and every experimental figure live on the project page, which is built from the same artefacts a pipeline run produces. No code, no model weights and no benchmark data are here yet.
@article{jiang2026vibebuilding,
title = {Vibe Building},
author = {Jiang, Yongqing and Luo, Haoran and Wang, Jianze and
Zhou, Xin and Dai, Kaoshan and Shen, Zhiqi},
journal = {arXiv preprint arXiv:2610.08156},
year = {2026},
doi = {10.48550/arXiv.2610.08156}
}The analysis core is built on OpenSeesPy. Geometry and rendering use Rhino and Blender. The held-out seismic library draws on the FEMA P-695 far-field set and the PEER NGA-West2 database.
Correspondence to yong.qingjiang@outlook.com.
Not yet determined. A license will be stated when the code and VB-Bench are released. Until then no license is granted for the contents of this repository.