Vibe Building

1 INTRODUCTION

• The Vibe Building task replaces vibe coding’s permissive “runs” criterion with whether a building design “stands” under code and structural constraints.Unlike software, a building cannot simply be reverted and regenerated after an error.Source: §1 · p. 1
• Prior agents either leave physics passive or let a language model render the verdict; the paper calls for deterministic evaluation under code-prescribed wind and seismic envelopes.The physics engine returns a margin for each constraint.Source: §1 · pp. 1–2
• Across four LLMs, held-out pass-rate variation was 3.3 pp for PE-Loop on seismic and 4.1 pp on wind, versus 23.8 and 27.7 pp for the best baseline per backbone.Figure 2a reports whiskers of ±1 SE.Source: Figure 2 · p. 2
• PE-Loop makes the physics engine the sole acceptance authority; its per-constraint margins guide section search and topology repair while the language model proposes discrete revisions.The loop terminates only when every strength and serviceability constraint holds.Source: §1 · pp. 2–3
• The paper formalizes multimodal code-compliant building design with in-loop compliance, held-out dynamic re-checks, and a constructability gate, and builds a benchmark of 3,577 instances across six code families.The benchmark includes 1,311 gradable tasks and a code knowledge base with limits traceable to clauses.Source: §1 · p. 2
• 58.43% of designs accepted by an LLM judge were false positives; PE-Loop had the best verified success under three of four backbones and the best held-out seismic pass rate under all four.The comparison replaces the physics verdict with an LLM judge.Source: §1 · p. 3

2 TASK: VIBE BUILDING

• An instance pairs a massing observation with a site that fixes its code family and gravity, wind, and seismic provisions; observations may be numeric text, sketch with text, or photographs alone.For sketch and image inputs, perception estimates the massing, but adjudication uses the ground-truth massing.Source: §2 · pp. 3–4
• Candidates combine discrete structural choices and gridded column size, and are evaluated against code-bounded demand quantities under gravity, wind, and seismic load cases.Static equilibrium and effective mass participation are zero-reference rejection gates before code limits are checked.Source: §2 · p. 3
• Each constraint returns a signed normalized margin that is nonnegative when satisfied; detailing clauses separately bound member proportions such as beam clear-span-to-depth ratio.The agent receives both the full margin vector and demand vector at every evaluation.Source: §2 · p. 3
• Episodes are bounded by a simulator-call budget and end with a returned design, budget exhaustion, or a scored ESCALATED outcome when available design levers are insufficient.The terminal reward is sparse feasibility, while the ideal return minimizes material among feasible realizations.Source: §2 · pp. 3–4
• Among feasible realizations, the ideal objective minimizes concrete volume, counting frames, cores, and walls.The feasible set contains designs satisfying every constraint.Source: §2 · pp. 3–4
• A returned design is certified with code-compliance margins, held-out dynamic checks under recorded ground motions and wind series, and compliance with the detailing set.The held-out seismic and wind series are not used for optimization.Source: §2 · p. 4

3 BENCHMARK: VB-BENCH

• The instance layer contains 3,577 buildings across four construction routes, while the task layer contains 1,311 gradable tasks stratified by solvability and intent, design, or modeling questions.The instance routes range from synthetic examples and literature-grade reconstructions to real anchors and complex-massing towers.Source: §3 · p. 4
• Every sample receives expert review, physical adjudication under gravity, wind, and seismic loads, and time-history verification; code limits are clause-traceable in a knowledge base compiled from 86 codes across six families.The task layer also includes 1,197 paired variants and a 372-task quarantine set against contamination.Source: §3 · p. 4

4 METHOD: PHYSICS-ENGINE-IN-THE-LOOP

• The deterministic margin oracle alone decides whether a design advances; language-model outputs change the next candidate, never the verdict.Constraints analytic in the design variables are enforced by construction, while simulator-dependent constraints are adjudicated by the oracle.Source: §4 · p. 4
• Physics process reward uses constraint margins to assess feasibility, while the violated set directs repairs.The loop consumes margin signs; magnitudes enter only the certificate.Source: §4 · p. 5
• The design varies square-column side b on a grid and treats core and wall counts, core geometry, and section choices as discrete.For sketch and image inputs, perception and intent calls set the initial core and wall counts; adjudication uses the true massing.Source: §4 · p. 5
• Under a monotone-margin assumption along the projected column axis, C2 finds the minimum feasible column side by probing, bisection, and confirmation.The search takes at most eight evaluations, and a returned design is re-verified.Source: §4 · p. 5
• When a round does not return PASS, the language model chooses a section-, topology-, or system-level action based on the last violated set.The method uses three outer rounds and a shared budget of 15 evaluations; the ultra tier adds a loop that accepts only feasible, strictly cheaper concrete-volume candidates.Source: §4 · pp. 5–6
• The Πbuild projection enforces the beam-detailing clause before evaluation by clipping beam depth and capping the maximum column side.The method states that every returned design is feasible even where its assumption fails; the Table 1 PE-Loop rows omit this projection.Source: §4 · p. 6
• Returned designs receive held-out seismic response-history and wind time-history checks, neither of which enters the loop.The certificate combines these held-out margins with detailing margins; replacing the in-loop verdict with an LLM judge measures the cost of changing acceptance authority.Source: §4 · p. 6

5 EXPERIMENTS

• The main split compares five systems across four LLMs on 130 topology-decision instances, with rates averaged over three seeds.Verified PASS requires static compliance, held-out time-history verification, and the constructability gate.Source: §5 · pp. 6–7

5.1 MAIN RESULTS

• PE-Loop leads Verified PASS in three of four blocks, held-out seismic in all four, and held-out wind in three.Under DeepSeek-V4-Flash, its 7.43 pp raw lead shrinks to 4.61 pp certified.Source: Figure 7 · p. 7, §5.1 · pp. 7–8

5.2 ABLATIONS

• Replacing the margin oracle with an LLM judge yields 49.44% claimed PASS but 20.56% physics-verified PASS, making 58.43% of acceptances false.A disjoint stronger judge still produces 56.94% false positives; the margin oracle's claimed and delivered PASS both equal 77.78%.Source: Figure 8 · p. 8, §5.2 · p. 8
• Adding Πbuild lowers raw PASS from 91.28 to 61.54% but raises Verified PASS from 35.64 to 45.38%.Concrete use falls from 5,610.1 to 5,032.0 m3; 123 of 126 constructability violations were repairable by clipping one section scale.Source: Figure 9 · p. 8, §5.2 · pp. 8–9

5.3 BUDGET AND WHAT THE LANGUAGE MODEL BUYS

• Against pure search on the 59-instance reference set, PE-Loop ties feasibility at 98.31% but improves NFE@pass and held-out wind pass rate.NFE@pass is 55.86% versus 97.78–100.00%, and wind pass is 72.41% versus 18.97–31.03%; concrete use is 5.66% higher than pure search.Source: Figure 10 · p. 9, §5.3 · p. 9
• PE-Loop's escalation action keeps budget exhaustion at or below 6.41% across LLMs, while baselines exhaust the budget on 58.19–91.81% of DeepSeek-V4-Flash answers.On identical seed-0 instances, success rises from 3.85% at light effort to 91.54% at high effort, then falls by 10.00 pp at ultra effort.Source: §5.3 · p. 9

6 CONCLUSION AND LIMITATIONS

• Reliability rests on who holds the verdict: replacing the deterministic engine with a judge turns most accepted designs into false positives, while certified success remains below one half.The held-out checks share the loop’s physical model; the authors identify independent-solver validation, broader clause coverage, and multi-coordinate search as next steps.Source: §6 · p. 9

AI USE STATEMENT

• The authors used generative AI to interpret results, edit text, create figures and scripts, and write verification code, then reviewed AI-assisted work and recomputed reported statistics.They state that AI was not used to generate synthetic data, implement PE-Loop or the evaluation harness, design experiments, or formulate hypotheses.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The designs are schematic and lack licensed-engineer review; certificates establish compliance with sourced clauses under stated analyses, not building safety..Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• The authors report that the physics oracle is deterministic, while language-model API access prevents bit-for-bit reproducibility; reported rates therefore use three seeds and standard errors.The planned release includes frozen result tables, per-episode records, seeds, model identifiers, and scripts to regenerate figures and re-assert reported numbers.Source: REPRODUCIBILITY STATEMENT · p. 10

A RELATED WORK

• Prior work spans LLM structural generation and checking, physics-guided layout generation, and tool-feedback agents, but the paper targets designs meeting multiple externally specified performance levels simultaneously.It extends a section-adjustment repair loop to topology-level repair and building-scale evaluation with held-out checks.Source: A RELATED WORK · pp. 16–17

• VB-Bench contains 3,577 buildings across four construction routes and 1,311 gradable tasks across intent, design, and modeling questions.It also includes 1,197 paired variants and a 372-task quarantine set against contamination.Source: B VB-BENCH DETAILS · p. 17
• Wind governs the majority of instances above 150 m; the code-source graph visualizes linked entries and marks two landmark pages.The connected component shown contains 583 of 625 entries; 42 unlinked or small-component pages are omitted.Source: Figure 11 · p. 18
• The 260-instance main split samples 130 topology-decision cases, 110 section-only controls, and 20 capability-boundary cases; only the topology-decision stratum separates systems.The pool follows screening and dynamic solvability calibration; 32 image-only cases are quarantined, and comparative statistics exclude the 20 boundary instances.Source: B VB-BENCH DETAILS · p. 18
• Acceptance requires expert review, physical adjudication, and time-history verification under gravity plus separate seismic and wind cases, using the most unfavorable response values.Acceptance limits scale with building height.Source: B VB-BENCH DETAILS · p. 18
• Held-out seismic selection uses only the 44 FEMA P-695 far-field records; held-out wind checks draw from eight field and eight spectrum-consistent synthetic series.The records are drawn from fixed public libraries; the wind series are ten minutes at 5 Hz.Source: B VB-BENCH DETAILS · pp. 18–19
• The code knowledge base organizes 86 original codes from six countries and regions into 625 cross-linked entries, with task limits traceable to clauses.It covers seismic and wind provisions, loads, concrete design rules, and code revision history.Source: B VB-BENCH DETAILS · p. 19
• Each compliance instance carries a code-family detailing set; the current release sources the beam clear-span-to-depth limit ln/hb ≥4 from two Chinese code clauses.The limit is evaluated geometrically from the returned member.Source: B VB-BENCH DETAILS · p. 19

• Four columns show the massing, rendering, OpenSeesPy frame idealization, and response; the first two depict what a designer sees and the last two what the physics engine analyzes.The response column shows peak roof drift under a linear time-history analysis, with members colored by displacement magnitude.Source: C ONE MASSING, FOUR PROJECTORS · p. 20
• The classic examples are three pre-existing Rhino massings converted into frames, while the three mega-frame towers came from the benchmark’s LLM-free search.Tower massings are rendered from lofted perimeter rings; the illustrated structures are not Section 5 episodes.Source: C ONE MASSING, FOUR PROJECTORS · p. 20
• The mega-frames are drift-certified only; their model omits top acceleration, period ratio, and shear-weight and stiffness-weight checks.The response model is screening-level and elastic, with no material nonlinearity, soil, or comparison against measured or published building response.Source: C ONE MASSING, FOUR PROJECTORS · p. 20
• The classic-building figure includes a slab tower, a slender tower with low wings, and a courtyard block; the town hall’s four separate pieces move independently.The six frame models’ sizes, periods, connected pieces, and peak responses are listed in Table 2.Source: Figure 13 · p. 21, Table 2 · p. 23

D.1 IN-LOOP CONSTRAINT SET AND THE ROLE OF THE PPR

• The in-loop set covers code-family checks such as drift, torsion, shear-to-weight, stiffness, and period ratios; tall buildings also receive wind-induced peak-acceleration checks.Recorded-ground-motion response-history analysis is held out of the in-loop constraint set.Source: §D.1 · p. 23
• The PPR guides repair using violated constraints before feasibility and ranks feasible candidates by concrete volume; search consumes margin signs, not magnitudes.The authors note that this use has no policy-invariance guarantee and leave magnitude-based action ordering as a possible extension.Source: §D.1 · p. 23

D.2 INITIAL TOPOLOGY AND THE COLUMN GRID

• Initial topology encodes system, governing load, and strategy in discrete design choices; the core geometry and section triple are drawn from finite menus.Defaults include a 0.15 core-to-floor area ratio, aspect ratio 1, zero offset, unit beam-section and concrete-grade scales, and zero taper.Source: §D.2 · p. 23
• The column grid uses 50 mm steps from 0.30 m to 2.50 m, yielding 45 candidate sizes and fixing the bisection cost in Lemma 1.The upper bound is shared across compared systems; the authors describe the grid as a convention, not a derivation.Source: §D.2 · pp. 23–24

D.3 DETAILING SET AND PROJECTION

• For a 6 m bay and ςbeam = 1, the clear-span-to-depth limit ln/hb ≥4 is violated when b exceeds 1.43 m, beginning at 1.45 m on the grid.The constraint binds for thick columns because beam depth grows while clear span shrinks as b increases.Source: §D.3 · p. 24
• Πbuild clips beam scale to ln/(3.2b) and caps column size at the largest grid value no greater than bay/2.6, enforcing the clear-span-to-depth clause before evaluation.The projection is analytic and costs no oracle calls; it follows from ln/hb ≥4 and the minimum beam-scale menu value of 0.5.Source: §D.3 · p. 24
• On the 59-instance search-sufficient reference set pooled over three seeds, Figure 15 measures clear-span-to-depth violations among physically feasible returned designs; all but three are repairable by clipping beam scale alone.The three exceptions have b > bay/2.6.Source: §D.3 · p. 24
• Two additional clause families are monitored but not sourced: the beam-width rule is not violated by returned designs, while other limits require unavailable or nonuniform inputs.The seismic-grade-dependent minimum column side needs a grade not defined uniformly across code families; the axial-compression-ratio limit needs member forces unavailable to the harness.Source: §D.3 · p. 24

D.4 PSEUDOCODE

• In the Πbuild variant, every candidate is projected before oracle evaluation; harness re-verification and final L4 verification are outside the agent.The pseudocode distinguishes language-model calls from charged margin-oracle calls.Source: §D.4 · p. 24

D.5 CORRECTNESS OF C2 AND ITS EVALUATION BOUND

• Every system returns unbuildable designs in 42.75–72.41% of cases; 123 of 126 violations for PE-Loop without Πbuild are repairable by clipping one section scale.Figure 15 reports the violating share among feasible designs in the 59-instance reference set.Source: Figure 15 · p. 25
• Under Assumption 1, feasible column sizes form an up-set, so C2 finds the minimum feasible size with at most 8 evaluations, or 9 including one outer-round re-verification.Without the assumption, a PASS still confirms feasibility; only minimality along the column-size axis may be lost.Source: §D.5 · p. 25
• C2 freezes the section triple during bisection because varying beam section, concrete grade, or taper would break monotonicity and expand the evaluation search.If the evaluation budget truncates the search, the final confirmation determines whether the returned column size is feasible.Source: §D.5 · p. 26

D.6 C1 ACTION SET AND THE BUDGETED OUTER LOOP

• When column resizing cannot find a feasible design at its upper bound, C1 proposes a section-, topology-, or system-level change based on violated constraints; exhaustion yields ESCALATED.Its actions include changing the section triple, adding cores or walls, adjusting core geometry, or revising the lateral system.Source: §D.6 · p. 27

D.7 MATERIAL LOOP C3

• In the ultra tier, C3 tests five section-triple candidates and accepts only feasible designs with strictly lower concrete volume, under a four-evaluation cap.Candidates are evaluated independently from the pre-C3 design, and the cheapest feasible candidate is adopted.Source: §D.7 · p. 27

D.8 COMPUTE TIERS

• The high tier is the main-table setting with B = 15 and K = 3, while success rate increases from light to high; C3 and held-out L4 verification run only at ultra and high/ultra, respectively.Light uses (B, K) = (5, 1), medium (10, 2), high (15, 3), and ultra (25, 4); only oracle calls count toward B.Source: §D.8 · pp. 27–28

D.9 TERMINATION SEMANTICS

• ESCALATED is a scored termination after levers or budget are exhausted, whereas FAIL denotes a crash; system-side non-delivery, timeout, or crash counts as infeasible.Harness cases that cannot be synthesized or lack a sourced constraint are excluded from the denominator.Source: §D.9 · p. 28

D.10 FINAL VERIFICATION L4

• Because seismic response-history analysis is one to two orders of magnitude more expensive than elastic static analysis, it is reserved for a final check rather than placed inside the iteration loop.L4 checks every returned design under recorded-ground-motion seismic response history and full wind time history; it shares the loop's physical model and cannot detect systematic model-form errors.Source: §D.10 · p. 28
• The seismic re-check uses one spectrum-matched, amplitude-scaled record per design, while the wind re-check uses three of sixteen ten-minute series rescaled to site turbulence intensity.The seismic check evaluates peak inter-story drift; wind checks drift and top-floor acceleration, with acceleration unevaluated for code families lacking a tabulated ten-year pressure.Source: §D.10 · p. 28

D.11 ARCHITECTURAL PLACEMENT

• The pipeline is an iterative task graph in which deterministic physics tools hold acceptance authority rather than a supervising model.Language-model calls and physics tools exchange feedback so the agent can choose whether to repair or escalate in each round.Source: §D.11 · p. 29

E ADDITIONAL EXPERIMENTAL RESULTS

• The 130 scored topology-decision instances are a stratified sample of 37.7% of that pool; section-only controls pass at 93.6–97.9% and boundary instances at none.Sampling is non-uniform by height: 87.0% of the 23 instances above 250 m are included, versus 4.8% of the 1,239 below 24 m.Source: E ADDITIONAL EXPERIMENTAL RESULTS · p. 29
• On the 60-instance grounding ablation, the judge received code limits and design parameters but no simulator quantities, and judge arms were not checked under held-out loading.The oracle arm's deterministic physics check was the lossless reference; the main-table PE-Loop pass rate on those 180 answers was 81.67%.Source: Table 5 · p. 31
• 9.74 pp: certified success rises with constructability projection despite a 29.74 pp raw PASS drop; with projection, Build% equals PASS%.The projection ablation is limited to the main split, one LLM, and topology-decision instances; the unprojected arm covers all 260 instances.Source: Table 6 · p. 31, E ADDITIONAL EXPERIMENTAL RESULTS · p. 31
• Under DeepSeek-V4-Flash, 66.6% of compliant designs pass held-out seismic checks, 76.7% pass wind checks, and 50.8% clear detailing.Of 126 PE-Loop detailing violations, 123 (97.6%) are repairable by clipping beam-depth scale alone.Source: E ADDITIONAL EXPERIMENTAL RESULTS · p. 31

E.1 THE SEARCH-SUFFICIENT REFERENCE SET: 59 INSTANCES

• Only Codex differs significantly from exhaustive search on the 59-instance reference set: search solves 18 instances Codex does not, and none in the reverse direction (Holm-corrected p = 1.14 × 10^-4).The other four systems have 0–2 discordant instances; p = 1 with no discordant pair indicates instance-by-instance agreement, not equivalence.Source: §E.1 · p. 32

E.2 EFFORT TIERS AND THE ANYTIME CONTRACT

• At effort B=25, ultra returns 81.54% PASS versus 91.54% (119 of 130) at B=15; extra rounds can revise settled topologies for the worse.These are seed-0 results on the same 130 topology-decision instances; wall-clock is not monotone across effort tiers because it includes gateway queueing.Source: §E.2 · p. 32

E.3 DENOMINATOR DISCLOSURE

• Rates use each arm's own answer denominator; GLM-block non-deliveries, timeouts, and agent-side unevaluated answers remain in it.The main protocol gives each system identical instances, margin oracle, and budget within an LLM block; baseline gateway waits vary substantially.Source: E ADDITIONAL EXPERIMENTAL RESULTS · pp. 29–30, §E.3 · p. 32

E.4 THREATS TO VALIDITY OF THESE NUMBERS

• Area-derived footprints assume a fixed 1.3 aspect ratio, biasing bay width and detailing margins for multi-block massings; the main split and reference set are not comparable.The projection ablation also covers only one LLM and 130 topology-decision instances, unlike its 260-instance unprojected arm.Source: E ADDITIONAL EXPERIMENTAL RESULTS · p. 31, §E.4 · p. 32