Physics-Engine-in-the-Loop

Vibe Building

A deterministic physics engine, not a language model, decides whether a building design is accepted.

Yongqing Jiang1†, Haoran Luo1*, Jianze Wang2, Xin Zhou1, Kaoshan Dai2& Zhiqi Shen1*

1Nanyang Technological University, Singapore 2Sichuan University, China

†Project leader. *Corresponding authors. Correspondence to yong.qingjiang@outlook.com.

Paper (PDF) · 18 MB arXiv · 2610.08156 Code (GitHub) Email

Abstract

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.

Vibe Building overview

The task

An instance pairs an observation with a site configuration. The observation may be a textual brief, a hand-drawn sketch, or a set of site photographs. The design must satisfy the code envelope for that site, and it is carried to a certified optimum rather than to a first feasible guess.

Figure 3: The definition and overview of the Vibe Building task: an instance pairs an observation (text, sketch or site
Figure 3. The definition and overview of the Vibe Building task: an instance pairs an observation (text, sketch or site photographs) with a site configuration, and the design is carried to a certified optimum.

Method

The language model proposes discrete moves and pays nothing for them. The physics engine decides, and every decision costs one evaluation of the budget. The loop terminates only when every strength and serviceability constraint holds.

Figure 5: Overview of PE-Loop. The meta level evolves discrete decisions (intent, system, strategy); the base level runs
Figure 5. Overview of PE-Loop. 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.
  1. B0–B2 language model

    Perception and intent

    One vision-language call turns a sketch or an image into a massing. A second call fixes the intent (the lateral system, the governing load, and the strategy) and with it the initial discrete state. Neither call is charged to the evaluation budget.

  2. C1 language model

    One discrete repair per round

    Given only the set of violated constraints, the model picks one action: adjust the section triple, add a core or a shear wall, change the core plan, or change the lateral system. Three rounds, one lever level each.

  3. C2 physics engine

    The minimal feasible column

    With the discrete state fixed, the column side is searched on a 45-value, 50 mm grid by an upper-bound probe, six bisection steps, and a confirmation. This costs at most 9 oracle calls per round, and the first passing side is also the least-material one.

  4. C3 physics engine

    Material loop, ultra tier only

    After a pass, a finite menu along the section triple is enumerated and a candidate is accepted only if it is still feasible and cheaper. The final design is re-measured, and the held-out analyses and the certificate are computed offline.

Results

Swap the backbone, keep the reliability

Because the verdict never depends on the proposer, the held-out pass rate barely moves when the language model behind it changes. A baseline that judges itself inherits the model’s variance instead.

3.3 pp PE-Loop · held-out seismic
4.1 pp PE-Loop · held-out wind
23.8 pp best baseline · seismic
27.7 pp best baseline · wind

Spread across the four backbones: DeepSeek-V4-Flash Qwen3.8-Flash GLM 5.2 Fast Kimi K3

Where certified success is won

Raw compliance and certified success are not the same ranking. Designs that pass in the loop still fall at the constructability gate and at the held-out dynamic checks, and systems reorder once those gates are applied.

Figure 6: Where certified success is won, and three interventions on PE-Loop. (a) Gate survival. (b) What a language-mod
Figure 6. Where certified success is won, and three interventions on PE-Loop. (a) Gate survival. (b) What a language-model judge claims versus what physics confirms. (c) What the constructability projection costs and pays. (d) The effort ladder.
Figure 7: Survival through the acceptance gates on all four LLM blocks.
Figure 7. Survival through the acceptance gates on all four LLM blocks.

Who holds the verdict

Replace the physics verdict with a language-model judge that sees the brief, the code limits and the design parameters but no simulator output. Everything else is fixed.

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

Table 5. The margin oracle claims exactly what it delivers. The language-model judge claims 49.44% and delivers 20.56%. The oracle’s two zeros are zero by construction, not for want of data: its gate is the physics check itself, so what it reports and what verification confirms are the same quantity, leaving no false acceptance and no evaluation spent on a misdirected repair.

Figure 8: Who holds the verdict. What each gate accepts, split into physics-confirmed and false, and what handing the ve
Figure 8. Who holds the verdict. What each gate accepts, split into physics-confirmed and false, and what handing the verdict to a model costs in wasted evaluations and wall-clock time.

Constructability costs raw compliance and pays certified success

With the constructability projection, every statically compliant design satisfies the detailing clause by construction, so the detailing gap closes to zero. Raw PASS drops by 29.74 pp and certified success rises by 9.74 pp.

Figure 9: The constructability projection trades raw compliance for certified compliance: coral is what it costs, blue w
Figure 9. The constructability projection trades raw compliance for certified compliance: coral is what it costs, blue what it pays.

What the language model actually buys

On the 59-instance reference set that pure search already solves, PE-Loop ties LLM-free search on feasibility at 98.31% with no discordant instances, so the language model buys no feasibility. What it buys is budget (55.86% NFE@pass against 97.78–100.00%) and held-out generalization (72.41% wind against 18.97–31.03%). The price is material: 5.66% more concrete than pure search.

VB-Bench

3,577 physics-adjudicated instances
6 code families
1,311 gradable tasks

Every code limit in the knowledge base is traceable to the clause it comes from.

Figure 2c: Instance locations by seismic code family. Marker size is building height; colour is the code family the insta
Figure 2c. Instance locations by seismic code family. Marker size is building height; colour is the code family the instance is adjudicated under.
Figure 4: VB-Bench overview. (a) Instance composition by route (inner) and code family (outer). (b) Height-slenderness o
Figure 4. VB-Bench overview. (a) Instance composition by route (inner) and code family (outer). (b) Height-slenderness of every instance. (c) Geographic coverage of the real instances.

GB EC8 ASCE 7 BSL IS NZS

One building, four stages

The same massing carried through the whole chain: the finished render, the Rhino massing being built, the OpenSeesPy frame the physics engine judges, and the Blender turntable. Every column is a real artefact of the run, not an illustration.

pipeline towerBurj Khalifa828.0 m · 20,736 elements
RhinoBurj Khalifa, Rhino view
OpenSeesPyBurj Khalifa, OpenSeesPy view
First modeBurj Khalifa, First mode view
BlenderBurj Khalifa, Blender view
pipeline towerDesign Shanghai Megaframe630.0 m · 17,964 elements
RhinoDesign Shanghai Megaframe, Rhino view
OpenSeesPyDesign Shanghai Megaframe, OpenSeesPy view
First modeDesign Shanghai Megaframe, First mode view
BlenderDesign Shanghai Megaframe, Blender view
classicWindmill Interlocking Campus18.0 m · 6 storeys · 1,128 elements
RhinoWindmill Interlocking Campus, Rhino view
OpenSeesPyWindmill Interlocking Campus, OpenSeesPy view
First modeWindmill Interlocking Campus, First mode view
BlenderWindmill Interlocking Campus, Blender view
classicHilversum Town Hall Redesign48.0 m · 13 storeys · 454 elements
RhinoHilversum Town Hall Redesign, Rhino view
OpenSeesPyHilversum Town Hall Redesign, OpenSeesPy view
First modeHilversum Town Hall Redesign, First mode view
BlenderHilversum Town Hall Redesign, Blender view

What the designs are checked against

Verification runs against held-out record libraries, not against the loads the design was tuned on: forty-four far-field ground motions and sixteen ten-minute wind series, eight recorded in the field and eight synthesised to match the Kaimal and Davenport targets.

Figure 12: The held-out record libraries. (a) Unscaled 5%-damped pseudo-acceleration response spectra of the 44 FEMA P-69
Figure 12. The held-out record libraries. (a) Unscaled 5%-damped pseudo-acceleration response spectra of the 44 FEMA P-695 far-field records, with their median and 16–84% band. (b) Normalised power spectra of the eight field-recorded wind series and the eight spectrum-consistent synthetic series, against the Kaimal and Davenport targets used for synthesis.

BibTeX

Preprint on arXiv. Replace with the proceedings entry after the decision.

@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}
}