MineCEraft: Evaluating Language Models as Construction Engineers in the World of Minecraft

# Evaluating Language Models as Construction Engineers in the World of Minecraft
Source: [https://arxiv.org/html/2608.28884](https://arxiv.org/html/2608.28884)
Sewoong LeeAffiliation:Siebel School of Computing and Data ScienceThe Grainger College of EngineeringUniversity of Illinois Urbana\-ChampaignRisham SidhuAffiliation:Siebel School of Computing and Data ScienceThe Grainger College of EngineeringUniversity of Illinois Urbana\-ChampaignJulia HockenmaierAffiliation:Siebel School of Computing and Data ScienceThe Grainger College of EngineeringUniversity of Illinois Urbana\-ChampaignYoonhwa JungAffiliation:Department of Civil and Coastal EngineeringUniversity of FloridaCorrespondence:[samuel27@illinois\.edu](mailto:[email protected]),[rsidhu3@illinois\.edu](mailto:[email protected]),[juliahmr@illinois\.edu](mailto:[email protected]),[yoonhwa\.jung@ufl\.edu](mailto:[email protected])

###### Abstract

We introduce MineCEraft \(Minecraft Construction Engineering Benchmark, pronouncedmine\-see\-ee\-raft\), an easy\-to\-use, open\-source111Our code is available at[https://github\.com/SewoongLee/mineCEraft](https://github.com/SewoongLee/mineCEraft)\.benchmark designed to systematically evaluate the reliability and limitations of LLMs for construction tasks in Minecraft\. The MineCEraft benchmark comprises 723 domain\-expert hand\-crafted natural\-language instructions with programmatically verifiable evaluation, spanning 17 distinct task categories, providing a safe and controllable experimental environment for assessing LLMs’ ability to perform realistic construction engineering tasks\. With this benchmark, we conduct an in\-depth evaluation of state\-of\-the\-art LLMs and perform a detailed error analysis, revealing key failure modes and practical challenges in applying LLMs to construction engineering tasks\.

![Refer to caption](https://arxiv.org/html/2608.28884v1/intro.png)Figure 1:LLM\-based agents fail to reliably execute even a simple instruction to build an arched bridge\.We decompose the task into rule\-based evaluation components\. This benchmark reveals the limitations that prevent LLM\-based agents from being reliably used in real\-world construction scenarios\.## 1Introduction

Construction is one of the oldest branches of engineering, accompanying humanity since its very origin\. While artificial intelligence has achieved remarkable progress in fields such as mathematics[Tao \(2024\)](https://arxiv.org/html/2608.28884#bib.bib21);[Chervonyi et al\. \(2025\)](https://arxiv.org/html/2608.28884#bib.bib11)and science\([Wang et al\., 2026](https://arxiv.org/html/2608.28884#bib.bib12)\), its limited contribution to construction engineering shows that Moravec’s paradox\([Moravec, 1988](https://arxiv.org/html/2608.28884#bib.bib6)\)still holds true in certain domains: what is easy for humans remains difficult for computers, and what is easy for computers remains difficult for humans\.

![Refer to caption](https://arxiv.org/html/2608.28884v1/ex-all.png)Figure 2:Examples of LLM\-generated constructions in MineCEraft tasks\.In the embodied Minecraft environment, the LLM frequently produces structurally flawed outputs even for relatively simple instructions\. All examples shown here were generated usingclaude\-3\-7\-sonnet\-20250219with the agent configuration of[White\* et al\. \(2025\)](https://arxiv.org/html/2608.28884#bib.bib16)\(Appendix[E\.2](https://arxiv.org/html/2608.28884#A5.SS2)\)\. Examples created by humans are provided in Appendix[H](https://arxiv.org/html/2608.28884#A8)\.While LLMs demonstrate impressive reasoning capabilities in many text\-based tasks, it remains an open question how well these skills translate to physical, spatial environments\. To systematically evaluate this, we abstract the core challenges of physical construction into a highly controlled, simplified environment\. A key question we investigate is:“How can we systematically demonstrate that even in a highly simplified environment like Minecraft, LLMs cannot reliably perform even basic construction tasks?”

We reveal the limitations of LLM\-based AI in reliably executing instructions, producing not only inaccuracies in shape and size, but also physically implausible and unstable structures that violate fundamental engineering constraints\. The contributions of this paper are as follows:

- •We formulate embodied construction as an engineering\-constrained task, where correctness requires physical plausibility, structural stability, and planning feasibility, enabling evaluation of structurally complex constructions beyond edit\-distance\-based metrics used in prior benchmarks\.
- •We introduce MineCEraft, an open\-source benchmark that combines natural language instruction following, long\-horizon embodied construction, and an evaluation pipeline that programmatically decomposes construction quality into accuracy, safety, and planning criteria\.
- •Using this framework, we provide systematic evidence that state\-of\-the\-art LLM\-based agents severely lack reliable spatial reasoning and planning capabilities, consistently failing to satisfy fundamental engineering\-level safety and stability constraints even in a simplified environment such as Minecraft\.

## 2Related Work

BenchmarkEmbodiedEnvironmentNatural Lang\.RequirementsLong\-horizonEvaluationPhysics\-awareConstructionEngineeringVoyager\([Wang et al\., 2024](https://arxiv.org/html/2608.28884#bib.bib5)\)✓✗✓✗✗EmbodiedBench\([Yang et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib20)\)✓✓✓✓✗MineCollab\([White\* et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib16)\)✓✗✓✗✗CEQuest[Wu et al\. \(2025\)](https://arxiv.org/html/2608.28884#bib.bib14)✗✓✗✗✓BAPv2[Jayannavar et al\. \(2026\)](https://arxiv.org/html/2608.28884#bib.bib17)✓✓✓✗✗MineCEraft \(Ours\)✓✓✓✓✓Table 1:Comparison of benchmarks\.### 2\.1Embodied AI

Embodied AI research studies how agents ground language in interactive environments, often focusing on navigation\([Thomason et al\., 2020](https://arxiv.org/html/2608.28884#bib.bib9);[Padmakumar et al\., 2022](https://arxiv.org/html/2608.28884#bib.bib7)\)or short\-horizon household manipulation\([Shridhar et al\., 2020](https://arxiv.org/html/2608.28884#bib.bib8)\)\. While these benchmarks evaluate sequential decision\-making, construction introduces a distinct challenge: agents must generate spatial structures that satisfy complex geometric, physical, and engineering constraints over long horizons\.

### 2\.2Minecraft\-based Construction and Language Grounding

Minecraft has been widely used as a platform for studying language\-guided construction, focusing both on instruction giving\([Narayan\-Chen et al\., 2019](https://arxiv.org/html/2608.28884#bib.bib19);[Köhn et al\., 2020](https://arxiv.org/html/2608.28884#bib.bib3)\)and following\([Gray et al\., 2019](https://arxiv.org/html/2608.28884#bib.bib2);[Jayannavar et al\., 2020](https://arxiv.org/html/2608.28884#bib.bib18)\)\. Recent works\([Chaturvedi et al\., 2024](https://arxiv.org/html/2608.28884#bib.bib10);[Kranti et al\., 2024](https://arxiv.org/html/2608.28884#bib.bib1);[Jayannavar et al\., 2026](https://arxiv.org/html/2608.28884#bib.bib17)\)have explored the use of text\-only LLMs for the Builder Action Prediction \(BAP\) task based on the Minecraft Dialogue Corpus \(MDC\) of[Narayan\-Chen et al\. \(2019\)](https://arxiv.org/html/2608.28884#bib.bib19)\. But the MineCEraft benchmark differs fundamentally from the MDC in that we are primarily interested in the construction tasks themselves, whereas the MDC focused on Minecraft as an interactive game, and back\-and\-forth dialogue between the instruction giver and follower \(builder\)\. MineCEraft does not assume that the builder can speak, but contains on more complex and physically realistic building tasks, disallowing the use of floating blocks common in MDC target structures, while allowing agents to use the full range of Minecraft materials \(not just a limited inventory of six colors\)\.

MineCollab\([White\* et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib16)\)introduces blueprint\-based construction tasks in Minecraft and evaluates agents’ ability to realize engineering\-style specifications\. However, its tasks focus on relatively basic structures \(e\.g\. pyramids\), and provide explicit coordinate\-level blueprints, implying that the primary challenge lies in execution rather than in interpreting natural\-language instructions\. In contrast, our MineCEraft evaluates realistic construction usingunderspecified instructions, rather than requiring an agent to replicate an exact target structure\.

### 2\.3Evaluation of Construction Engineering

Outside the Minecraft domain, several benchmarks have studied construction engineering knowledge in language models\. CEQuest\([Wu et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib14)\)evaluates large language models on construction engineering questions\. While this benchmark provides a useful evaluation of domain knowledge, it operates entirely in a text\-based setting and does not assess agents in embodied environments\.

Complementary work has explored physics\-based simulation within Minecraft\. For example,[Beck et al\. \(2024\)](https://arxiv.org/html/2608.28884#bib.bib13)introduced an elasticity solver for Minecraft structures that enables stress and deformation analysis of constructed buildings\. Although this work focuses on physical simulation rather than on language\-guided tasks, it demonstrates the potential to evaluate the structural properties of constructions in Minecraft environments\.

Taken together, existing work evaluates either language grounding in construction settings or engineering knowledge in text\-based environments\. However, none jointly evaluate natural\-language instruction following, embodied construction, long\-horizon spatial planning, and engineering\-aware structural evaluation\.

MineCEraft addresses this gap by introducing a benchmark for natural\-language\-guided construction tasks that incorporates engineering\-inspired evaluation criteria: it evaluates agents on their ability to interpret construction specifications, plan multi\-step building processes, and produce structurally valid constructions within a unified framework \(see Table[1](https://arxiv.org/html/2608.28884#S2.T1)\)\.

## 3Task Setup

The MineCEraft benchmark contains 723 natural\-language construction instructions hand\-crafted by a domain\-expert\. These tasks span 17 distinct task types designed to cover diverse construction scenarios and engineering requirements\.

### 3\.1Underspecified Instructions

A defining feature of MineCEraft is its use of underspecified instructions to evaluate this spatial reasoning\. Unlike traditional benchmarks that compare outputs against a single ground\-truth target, our tasks mirror real\-world construction by requiring agents to satisfy a set of open\-ended constraints\.

In our benchmark, a prompt such as“Build the frame \(including stairs\) of a two\-story building using concrete\. Each upper floor should be 4 blocks above the one below it\.”does not have one exact correct voxel configuration\. Instead, agents are evaluated on whether the final output meets explicitly defined criteria in Section[4](https://arxiv.org/html/2608.28884#S4)\(e\.g\., correct shape, size, material usage, physical plausibility\)\. Furthermore, tasks frequently involve initial instructions followed by multi\-turn revisions\. This tests the agent’s ability to adapt its spatial understanding to an evolving environment\.

![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/unsafe_arch.png)\(a\)An example ofphysically implausiblefloating blocks \(highlighted in red box\)\.
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/bridge_stress.png)\(b\)von Mises stress of an example of aflat\-shaped bridge with a span and height of 5 blocks \(max\.≈\\approx7\.8kPa\)\.
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/arch_stress.png)\(c\)von Mises stress of an example of anarch\-shaped bridge with a span and height of 5 blocks \(max\.≈\\approx5\.3kPa\)\.

Figure 3:Evaluating arch\-like structures beyond simple geometry\.Prompt:*“Build an arch bridge\.”*Physical plausibility is verified via breadth\-first search, penalizing disconnected floating blocks \(e\.g\., Figure[3\(a\)](https://arxiv.org/html/2608.28884#S3.F3.sf1), generated byclaude\-3\-7\-sonnet\-20250219\)\. Structural stability is measured via von Mises stress\. We use a standard flat bridge \(Figure[3\(b\)](https://arxiv.org/html/2608.28884#S3.F3.sf2),≈7\.8\\approx 7\.8kPa\) as the max threshold \(σmax\\sigma\_\{\\mathrm\{max\}\}\)\. In structural engineering, a true arch distributes loads more efficiently \(Figure[3\(c\)](https://arxiv.org/html/2608.28884#S3.F3.sf3),≈5\.3\\approx 5\.3kPa\), yielding a lower maximum stress than a flat bridge of the same span\. Any suboptimal arch exceeding the flat baseline stress \(σmax\>σref\\sigma\_\{\\max\}\>\\sigma\_\{\\mathrm\{ref\}\}\) receives a proportionally penalized stability score\.
### 3\.2Task Design and Coverage

MineCEraft includes a set of construction tasks that aim \(1\) to ensure broad diversity in construction scenarios, \(2\) to enable objective and precise evaluation through explicit rules or physics\-based computation, and \(3\) to reflect requirements that are realistic and plausible in real\-world construction practice\. Example error cases for qualitative analysis are shown in Figure[2](https://arxiv.org/html/2608.28884#S1.F2)\.

#### 3\.2\.1Structural Elements

This category covers fundamental structural components of building:foundations, walls, roofs, columns, andframes\. Tasks specify concrete requirements such as size, thickness, height, or number of layers\. Because these elements are defined by explicit geometric and material constraints, they allow straightforward evaluation using rule\-based checks on dimensions and connectivity\. They serve as basic tests of structural correctness before introducing more complex construction scenarios\.

#### 3\.2\.2Building Types

Beyond isolated structural elements, the Buildings category evaluates the construction of integrated buildings\. These tasks combine multiple components and introduce higher\-level requirements, including room layout, material constraints, accessibility, and planning over time\. The category includes a range of scenarios:

##### Basic Buildings

Instructions specify material assignments to different parts of a house, such as pillars, walls, roofs, and foundations\. Variants include different material combinations and prompt phrasings\. The illustrative example is shown in Figure[2](https://arxiv.org/html/2608.28884#S1.F2)\-B, and these tasks are evaluated using Algorithms[3](https://arxiv.org/html/2608.28884#alg3)and[11](https://arxiv.org/html/2608.28884#alg11)\.

![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/biased_pillar.png)\(a\)Biasedcolumn placement where the center column is positioned adjacent to another column\.
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/biased_pillar_stress.png)\(b\)von Mises stress underbiasedsupport \(max\.≈140​MPa\\approx 140\\,\\mathrm\{MPa\}\)\.
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/balanced_pillar_stress.png)\(c\)von Mises stress underbalancedsupport \(max\.≈60​MPa\\approx 60\\,\\mathrm\{MPa\}\)\.

Figure 4:MineCEraft shows that the LLM\-generated structures in \(a\) and \(b\) are structurally unstable compared to the alternative in \(c\)\.When the center column is attached to another column, the load is not distributed uniformly, leading to a biased stress field and a higher peak von Mises stress \(≈140​MPa\\approx 140\\,\\mathrm\{MPa\}\) compared to the balanced five\-column configuration \(≈60​MPa\\approx 60\\,\\mathrm\{MPa\}\)\. The construction was generated usingclaude\-3\-7\-sonnet\-20250219with the prompt:*“Build a 10×\\times10 flat roof supported by exactly five 2×\\times2×\\times2 columns\. Arrange the columns so that the roof is as structurally stable as possible\.”*
##### Multiple Rooms/Bedrooms

Tasks require constructing houses with a specified number of rooms \(e\.g\., one to five; see Algorithm[5](https://arxiv.org/html/2608.28884#alg5)\)\. The multiple bedroom instructions extend room\-based tasks by additionally specifying the number of beds to be included in the house using Algorithm[2](https://arxiv.org/html/2608.28884#alg2)\.

##### Multi\-story Buildings

Agents are instructed to construct upper floors or structural frames with explicit height and column constraints using Algorithm[8](https://arxiv.org/html/2608.28884#alg8)\.

#### 3\.2\.3Architectural Elements

##### Accessibility

Some tasks introduce functional requirements, such as wheelchair accessibility, single\-level layouts, minimum doorway width, or clearly defined exit paths \(Algorithm[7](https://arxiv.org/html/2608.28884#alg7)\)\.

##### Creative Requests

Creative variants impose abstract geometric properties, such as asymmetric structures or multiple ceiling heights\. Asymmetry is verified using the symmetry test described in Section 4\.1 \(Shape iii\), while ceiling requirements ensure that the structure reaches a minimum height without unsupported floating roofs \(Figure[2](https://arxiv.org/html/2608.28884#S1.F2)\-C\)\.

#### 3\.2\.4Civil Engineering Structures

A natural extension beyond building\-scale constructions is to consider civil engineering structures such as bridges\.

##### Bridges

Bridge tasks specify river widths and clearance requirements for boats of given dimensions\. At ground level, the structure must form two distinct support clusters separated by at least the required span \(Algorithm[6](https://arxiv.org/html/2608.28884#alg6)\)\.

##### Arch Bridges

We further introduce arch bridge variants that require curved upper and lower surfaces with explicit span and height constraints\. These tasks extend the Arch Bridge Challenge described in Appendix[B](https://arxiv.org/html/2608.28884#A2), which presents a simplified two\-dimensional diagnostic example, into a full three\-dimensional evaluation \(Figure[3](https://arxiv.org/html/2608.28884#S3.F3)\)\.

##### Domes

To extend curvature beyond bridges, we include dome\-like structures defined by base diameter and minimum height constraints, evaluated by Algorithms[4](https://arxiv.org/html/2608.28884#alg4)and[6](https://arxiv.org/html/2608.28884#alg6)In addition to instructions that explicitly use the term “dome,” we also include variants such as “igloo,” which differ in wording or material but impose essentially the same geometric requirements\.

#### 3\.2\.5Construction Management

##### Resource Optimization\.

Certain instructions limit available materials and require constructing the largest possible structure under the given constraint using Algorithms[2](https://arxiv.org/html/2608.28884#alg2)and[10](https://arxiv.org/html/2608.28884#alg10)\.

##### Single\-turn Planning\.

In these tasks, all construction requirements are provided in a single instruction, without specifying the order of execution\. The agent must determine an appropriate construction sequence as in Section[4\.3](https://arxiv.org/html/2608.28884#S4.SS3), which is evaluated by Algorithm[13](https://arxiv.org/html/2608.28884#alg13)\.

##### Multi\-turn Planning & Revision\.

Multi\-turn scenarios involve sequential modifications to an existing structure\. After each instruction that updates the required number of rooms, the resulting structure is validated using Algorithm[5](https://arxiv.org/html/2608.28884#alg5)\.

### 3\.3MineCEraft\-Lite: A Cost\-Efficient Subset

Since running the full benchmark can be computationally expensive, we also provide MineCEraft\-Lite, a stratified subset of tasks for cost\-efficient experiments\. This version is constructed via proportional sampling with a random seed across task categories and planning complexity levels to preserve the diversity and difficulty distribution of the full benchmark\. Specifically, MineCEraft\-Lite contains 25 tasks sampled from the 17 task categories of the full benchmark\. We recommend reporting results on the full benchmark whenever feasible, while MineCEraft\-Lite serves as a practical alternative for rapid prototyping and resource\-constrained settings, including human evaluation\. The representativeness of the lite version is discussed in Appendix[F](https://arxiv.org/html/2608.28884#A6)\.

## 4Evaluation Criteria

In this work, we design an evaluation pipeline that is fully verifiable with rule\-based algorithms\. As illustrated in Figure[2](https://arxiv.org/html/2608.28884#S1.F2), structural properties such as ground connectivity or span separation can be verified through explicit algorithms \(e\.g\., breadth\-first search for connectivity checks\)\.

This rule\-based approach provides another advantage\. Since correctness is defined by executable code rather than explicit ground\-truth text answers exposed to web\-crawling, the benchmark is inherently resistant to data contamination from LLM pre\-training, while simultaneously providing a ready\-made reward function for reinforcement learning applications\.

By decomposing each construction task into objectively verifiable constraints, we transform open\-ended instruction following into a set of well\-defined engineering criteria\. Specifically, we evaluate construction quality along three dimensions: accuracy, safety, and planning\.

### 4\.1Accuracy

Accuracy is evaluated along three dimensions,x,y,zx,y,z, using the set of blocks placed by the agent\. We represent the constructed structure as a block setℬ=\{\(x,y,z,material\)\}\\mathcal\{B\}=\\\{\(x,y,z,\\text\{material\}\)\\\}, where each tuple denotes the spatial coordinates and material type of a block\. In all cases, evaluation mapsℬ\\mathcal\{B\}together with task parameters \(e\.g\. expected counts, materials, heights\) to pass or fail\.

##### Material\.

Block types and counts are checked against the spec by matching each block’s material \(exact or substring\), optionally restricting to a subset by position \(e\.g\. corners, boundaries at a heightyy, or the top layer\), then requiring either a count condition\#​\{matching\}=k\\\#\\\{\\text\{matching\}\\\}=kfor a specified countkk, or a ratio condition\#​\{matching\}≥θ⋅\|subset\|\\\#\\\{\\text\{matching\}\\\}\\geq\\theta\\cdot\|\\text\{subset\}\|for a ratio thresholdθ∈\[0,1\]\\theta\\in\[0,1\]\.

##### Shape\.

Geometric constraints are verified in three ways: \(i\) connectivity—at a givenyy, clusters on the\(x,z\)\(x,z\)\-plane \(4\- or 8\-neighbors\) are formed, then cluster count, inter\-cluster distance, or radius/diameter are constrained, and BFS is used for reachability \(exit, rooms, passable doors\); \(ii\) surface concavity—for the top or bottom surface\(x,z\)↦y\(x,z\)\\mapsto y, the actualyymust lie above or below the linear interpolation between any two surface points; \(iii\) symmetry—whether the structure exactly matches itself under reflection or folding\.

##### Size\.

Dimensions are evaluated by computing axis rangesΔ​x=max⁡x−min⁡x\\Delta x=\\max x\-\\min x\(and similarlyΔ​y\\Delta y,Δ​z\\Delta z\) andmin⁡y\\min y,max⁡y\\max yoverℬ\\mathcal\{B\}, then comparing them to the specification\. The horizontal axesxxandzzare evaluated interchangeably, since the Minecraft coordinate system only uniquely distinguishes the vertical \(gravity\-aligned\) axisyyfrom the rotationally symmetricxx–zzground plane\.

### 4\.2Safety

In general, safety at construction sites is influenced by a wide range of factors, including regulatory compliance, hazard identification and assessment, and excavation processes\. Our study selects a subset of these factors that can be naturally and effectively observed within a simulation environment such as Minecraft, and automates their verification\.

##### Physical Plausibility Analysis

Typically, attempts to place blocks in mid\-air or to build from top to bottom rather than from bottom to top are common feasibility issues\. For example, as in Figure[3\(a\)](https://arxiv.org/html/2608.28884#S3.F3.sf1), whether a structure shows this issue can be checked by a breadth\-first search that verifies ground connectivity \(i\.e\., that every block is reachable from at least one block with y≤\\leq0 via face adjacency\) and converted into a 0/1 score for each prompt\.

##### Structural Stability Analysis

Structures with missing or unevenly distributed supports \(e\.g\., a wide roof on few columns\) are at risk of collapse\. We quantify this using the von Mises stress, following[Beck et al\. \(2024\)](https://arxiv.org/html/2608.28884#bib.bib13)\. For each construction, we run a simulation to obtain the maximum von Mises stressσmax\\sigma\_\{\\max\}over all blocks\. We compare this to a task\-specific optimal baseline,σref\\sigma\_\{\\mathrm\{ref\}\}, which represents the maximum stress of an ideal, human\-designed configuration for that specific task \(e\.g\., Figure[4\(c\)](https://arxiv.org/html/2608.28884#S3.F4.sf3): an evenly distributed column layout to support a roof\)\. The structural stability score is then computed asmin⁡\(1,σref/σmax\)\\min\\bigl\(1,\\;\\sigma\_\{\\mathrm\{ref\}\}/\\sigma\_\{\\max\}\\bigr\)\. Consequently, if the agent’s structure yields a maximum stress higher than the optimal reference \(σmax\>σref\\sigma\_\{\\max\}\>\\sigma\_\{\\mathrm\{ref\}\}\), its score is proportionally penalized to reflect suboptimal stability\. The detailed formulation is provided in Appendix[G](https://arxiv.org/html/2608.28884#A7)\.

CategorySubcategoryModelsHumanllama\-4 \-17b\-16eclaude\-4\-5 \-sonnetgpt\-5 \-minigemini\-3 \-proAccuracyMaterial27\.0%84\.5%92\.7%70\.4%97\.8%Shape22\.2%55\.8%53\.4%50\.3%99\.2%Size3\.2%68\.1%77\.0%52\.8%91\.7%SafetyPhysical Plausibility77\.8%85\.0%68\.3%92\.0%98\.3%Structural Stability25\.0%57\.1%50\.7%42\.7%93\.9%PlanningEfficiency47\.1%56\.2%55\.6%59\.7%95\.6%Dependency7\.4%96\.3%68\.6%78\.2%90\.0%Table 2:Performance on MineCEraft across different model families and sizes\.Human results are reported on MineCEraft\-Lite for reference \(see Appendix[E\.1](https://arxiv.org/html/2608.28884#A5.SS1)\)\. Overall, LLMs underperform human participants across most evaluation categories\. For Lite\-only comparisons across all models, see Table[10](https://arxiv.org/html/2608.28884#A6.T10)in Appendix[F](https://arxiv.org/html/2608.28884#A6)\. For evaluation results on more models, see Table[11](https://arxiv.org/html/2608.28884#A6.T11)in Appendix[F](https://arxiv.org/html/2608.28884#A6)\.

### 4\.3Planning

Construction inherently involves long\-horizon planning\. When placing tens to hundreds of blocks, organizing a sequence of actions over time determines what can be built and how efficiently the construction can proceed\. In real\-world construction practice, this is examined through scheduling reviews, which involve checking dependencies and measuring construction inefficiencies\. In this work, we evaluate planning along two complementary dimensions: dependency and efficiency\.

##### Dependency Analysis

One of the most critical aspects of construction scheduling is planning dependency\. For a simple example, a roof cannot be constructed before the supporting columns\. Such cases are detected by verifying the presence of required supporting materials before dependent components are placed\. In many tasks, each building element is explicitly assigned a distinct material\. This allows us to infer the construction order by tracking when blocks of each material are placed in the action log\. Based on this reconstruction, we determine whether structural elements were built in a valid sequence\. Specifically, the load\-bearing hierarchy, such as foundation→\\rightarrowcolumns→\\rightarrowwalls→\\rightarrowceiling/roof, must follow this order\. Violations are evaluated as failures\. In contrast, non\-structural tasks such as ceiling vs\. drywall or paint vs\. casework do not require a single correct ordering\. Such steps may be executed in either order without constituting an error\.

##### Efficiency Analysis

Another key planning challenge in construction is optimizing motion and time efficiency, often driven by movement\. For example, when human workers are instructed to lay bricks in a 10\-by\-10 pattern, they naturally use an L\-shaped traversal \(or lawnmower pattern\) that minimizes unnecessary movement\. In contrast, we observe that AI agents, particularly when actions are planned through code using a naive for\-loop structure, tend to place bricks sequentially from \(1,1\) to \(1,10\), and then return all the way back to \(2,1\) to continue placement\. We represent the ordered block placement sequence asℬ=\(b1,b2,…,bn\)\\mathcal\{B\}=\(b\_\{1\},b\_\{2\},\\ldots,b\_\{n\}\), where eachbi∈ℤ3b\_\{i\}\\in\\mathbb\{Z\}^\{3\}denotes the spatial coordinate of a placed block\. The total L1 \(Manhattan\) path length is then defined asL1total=∑i=1n−1‖bi\+1−bi‖1\.L\_\{1\}^\{\\mathrm\{total\}\}=\\sum\_\{i=1\}^\{n\-1\}\\\|b\_\{i\+1\}\-b\_\{i\}\\\|\_\{1\}\.The minimum possible length, when every step is to an adjacent cell, isn−1n\-1\. The efficiency score is defined asmin⁡\(1,\(n−1\)/L1total\)∈\[0,1\],\\min\\bigl\(1,\\;\(n\-1\)/L\_\{1\}^\{\\mathrm\{total\}\}\\bigr\)\\in\[0,1\],so that the optimal path yields11, and any extra distance or long jumps reduces the score\.

## 5Experiments

### 5\.1Agent Setup

We adoptmindcraft\([White\* et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib16)\)as our primary agent architecture\. Our benchmark is compatible withanyagent capable of \(i\) receiving natural language instructions as input and \(ii\) interfacing with Minecraft via the Mineflayer API\([PrismarineJS,](https://arxiv.org/html/2608.28884#bib.bib15)\)\. To our knowledge,mindcraft\([White\* et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib16)\)is among the few non\-finetuned LLM\-based agents evaluated on construction tasks capable of translating natural language reasoning into executable actions within a live Minecraft environment\.

At a high level, the agent consists of three components: \(a\) a language and planning module, \(b\) a coding interface to the Minecraft environment, and \(c\) a predefined skill library\. Given a natural language instruction, the LLM first produces a high\-level plan describing the construction procedure\. This plan is then translated by the coding component into executable JavaScript commands that interact with the environment through the Mineflayer API, using the provided skill library as building blocks for common actions\. The execution environment records detailed action logs \(Figure[1](https://arxiv.org/html/2608.28884#S0.F1)\); we parse these logs to recover the sequence of block placements and removals\. The resulting structures are then automatically graded using our rule\-based evaluation criteria\.

To ground the model in real\-world construction constraints, we modify its system prompt to explicitly encode engineering assumptions \(see Appendix[E\.2](https://arxiv.org/html/2608.28884#A5.SS2)for prompt details and the full skill library\)\. All other settings remain identical tomindcraft\.

llama\-4\-17b\-16eclaude\-4\-5\-sonnetgpt5\-minigemini\-3\-pro\#Problems\#CorrectOverall Acc\.\#CorrectOverall Acc\.\#CorrectOverall Acc\.\#CorrectOverall Acc\.72330\.4%19226\.6%28639\.6%27738\.3%

Table 3:Overall problem\-level accuracy across the 723 MineCEraft prompts\.Scores are computed by requiring that all success/failure evaluation criteria for a problem be satisfied simultaneously\. A more detailed instruction\-level aggregation method is explained in Appendix[D\.1](https://arxiv.org/html/2608.28884#A4.SS1)\. Even the strongest models achieve less than 40% overall accuracy\. For comparison, human participants achieve 90% accuracy on MineCEraft\-Lite\.
### 5\.2Experimental Results

##### Reliable construction remains challenging for LLMs\.

Table[2](https://arxiv.org/html/2608.28884#S4.T2)reports model performance on individual evaluation categories, including accuracy, safety, and planning metrics\. Although state\-of\-the\-art models achieve moderate scores on several of these criteria, solving an entire construction problem requires satisfying all applicable constraints simultaneously\.

We therefore evaluate problem\-level success across the 723 MineCEraft problems by checking whether a construction satisfies all evaluation criteria at once\. Under this strict definition,no model achieves an overall accuracy above 40%\(Table[3](https://arxiv.org/html/2608.28884#S5.T3)\), meaning that even the strongest models fail to correctly complete the majority of construction tasks\.

For comparison, when the same tasks were evaluated by two human participants with graduate\-level degrees in construction engineering, each provided with the same information as the LLM agents \(see Appendix[E\.1](https://arxiv.org/html/2608.28884#A5.SS1)for details\),human participants successfully solved roughly 90% of the problems\.This stark gap indicates that current LLMs remain far from the level of reliability required for construction tasks\.

##### ‘Step\-by\-step’ does*not*improve performance in construction\.

In domains such as mathematics, adding prompts like“Let’s think step by step\.”has been shown to improve performance by eliciting chain\-of\-thought reasoning\([Kojima et al\., 2022](https://arxiv.org/html/2608.28884#bib.bib4)\)\. However, in embodied construction tasks, adding this phrase did not lead to measurable performance gains\. Instead, performance improvements were observed only when prompts explicitly targeted construction\-specific constraints\. For example, adding“Let’s think about the order in which we should place the blocks to optimize the movement path\.”reduced inefficient block placement trajectories \(see Appendix[D](https://arxiv.org/html/2608.28884#A4)\)\. These findings suggest that improvements in verbal reasoning prompts do not automatically translate into physical environments\.

##### Robust to technical terms, sensitive to everyday verbs\.

For example, a field\-specific jargon such as SOMD \(slab on metal deck\) did not yield lower scores than simplified explanations, suggesting that construction\-specific vocabulary is reasonably well captured during pre\-training\. In contrast, performance often varied with the everyday verb used in the instruction\. For certain models, such as claude\-4\-5\-sonnet, tasks that differed only in the verb “build,” “create,” or “construct” produced different results, with “construct” leading to higher material and shape accuracy\. Detailed results are provided in Appendix[D](https://arxiv.org/html/2608.28884#A4)\.

##### LLMs are good at math, weaker at physics\.

While our results reveal substantial weaknesses of LLMs in physics\-grounded construction and planning, we observe an interesting counterpoint: when spatial reasoning can be reduced to a closed\-form, formula\-driven optimization problem, LLMs may outperform human participants\. For example, given the instruction:“With only 56 oak wood planks, construct the largest possible single\-room cabin\.”, an optimal solution requires recognizing that a hollow cubic structure with43−23=564^\{3\}\-2^\{3\}=56\. LLMs often effortlessly identify this volumetric relationship and construct the optimal structure\. In contrast,*none of*the human participants successfully discovered the optimal structure even after a number of trials and errors\.

## 6Conclusion

In this work, we introduced MineCEraft, an open\-source benchmark for evaluating large language models as construction engineers in a controlled Minecraft environment\. MineCEraft jointly assesses natural\-language instruction following, construction accuracy, safety, and long\-horizon planning using rule\-based and physically grounded metrics\. Our evaluation reveals that state\-of\-the\-art LLMs consistently fail to produce reliable and safe construction plans, even for seemingly simple tasks\. These results highlight fundamental limitations of current embodied language models in construction engineering and underscore the need for more robust reasoning, planning, and physical grounding\.

## Limitations

As the first benchmark to jointly evaluate natural\-language instruction following, long\-horizon construction, and physics\-aware structural stability in Minecraft, MineCEraft inevitably has several limitations\.

First, our evaluation focuses on detecting clear and objective violations of task requirements rather than sufficiently certifying construction quality\. To be specific, a construction that satisfies all rubric criteria cannot be guaranteed to be perfect in a real\-world setting; however, failure to achieve a full score indicates a clear violation of the specified requirements\. For example, when evaluating structural stability, identifying the globally optimal roof configuration is an NP\-hard combinatorial optimization problem\. Therefore, we adopt a reference\-based comparison, such as the balanced configuration shown in Figure[4\(c\)](https://arxiv.org/html/2608.28884#S3.F4.sf3)\. While current LLM performance makes such faults relatively easy to detect, more refined benchmarks will be necessary as model capabilities improve\.

Second, although Minecraft provides a highly flexible and controllable simulation environment, it remains an abstraction of the real world\. Although arch bridges can be constructed, certain structural types, such as suspension bridges, cannot be represented within the block\-based mechanics\. Additionally, our stress analysis follows the elasticity formulation of[Beck et al\. \(2024\)](https://arxiv.org/html/2608.28884#bib.bib13), and exploring alternative physical assumptions or more realistic material models constitutes an important direction for future work\.

Third, due to computational constraints, our experiments cover a limited set of state\-of\-the\-art LLMs and focus on text\-based interaction through Mineflayer\. This design enables efficient and reproducible evaluation with relatively modest computational resources compared to vision\-language models\. However, the absence of visual feedback may limit the ability of multimodal agents to fully leverage their visual grounding capabilities in our current setting\. Evaluating agents with egocentric visual perception and multimodal reasoning capabilities therefore remains an important direction for future research\.

Finally, our results are also sensitive to the choice of system prompt\. As shown in Appendix[D\.4](https://arxiv.org/html/2608.28884#A4.SS4), removing the brevity constraint can substantially change performance on several evaluation criteria, particularly dependency planning\. A systematic study of system\-prompt design and optimization is required in future work\.

## Acknowledgments

We sincerely thank Prashant Jayannavar and Siddarth Madala for their valuable feedback and helpful discussions\.

## References

- Z\. Beck, B\. Alpert, A\. Bowman, W\. R\. Watson, and A\. B\. TepoleElasticity solver in Minecraft for learning mechanics of materials by gaming\.Biomedical Engineering Education4\(1\),pp\. 129–135\.Cited by:[Appendix G](https://arxiv.org/html/2608.28884#A7.p1.1),[§2\.3](https://arxiv.org/html/2608.28884#S2.SS3.p2.1),[§4\.2](https://arxiv.org/html/2608.28884#S4.SS2.SSS0.Px2.p1.1),[Limitations](https://arxiv.org/html/2608.28884#Sx1.p3.1)\.
- Chaturvediet al\.\(2024\)A\. Chaturvedi, K\. Thompson, and N\. AsherNebula: a discourse aware Minecraft builder\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 6431–6443\.External Links:[Link](https://aclanthology.org/2024.findings-emnlp.374/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.374)Cited by:[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p1.1)\.
- Chervonyiet al\.\(2025\)Y\. Chervonyi, T\. H\. Trinh, M\. Olšák, X\. Yang, H\. H\. Nguyen, M\. Menegali, J\. Jung, J\. Kim, V\. Verma, Q\. V\. Le,et al\.Gold\-medalist performance in solving olympiad geometry with AlphaGeometry2\.Journal of Machine Learning Research26\(241\),pp\. 1–39\.Cited by:[§1](https://arxiv.org/html/2608.28884#S1.p1.1)\.
- Grayet al\.\(2019\)J\. Gray, K\. Srinet, Y\. Jernite, H\. Yu, Z\. Chen, D\. Guo, S\. Goyal, C\. L\. Zitnick, and A\. SzlamCraftAssist: a framework for dialogue\-enabled interactive agents\.External Links:1907\.08584,[Link](https://arxiv.org/abs/1907.08584)Cited by:[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p1.1)\.
- Jayannavaret al\.\(2020\)P\. Jayannavar, A\. Narayan\-Chen, and J\. HockenmaierLearning to execute instructions in a Minecraft dialogue\.InProceedings of the 58th annual meeting of the association for computational linguistics,pp\. 2589–2602\.Cited by:[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p1.1)\.
- Jayannavaret al\.\(2026\)P\. Jayannavar, L\. Ren, M\. Hudspeth, R\. Sidhu, C\. Lambert, A\. Cordes, E\. Kaplan, A\. Narayan\-Chen, and J\. HockenmaierBAP v2: an enhanced task framework for instruction following in Minecraft dialogues\.Computational Linguistics,pp\. 1–57\.Cited by:[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p1.1),[Table 1](https://arxiv.org/html/2608.28884#S2.T1.2.6.1)\.
- Köhnet al\.\(2020\)A\. Köhn, J\. Wichlacz, C\. Schäfer, Á\. Torralba, J\. Hoffmann, and A\. KollerMC\-Saar\-Instruct: a platform for Minecraft instruction giving agents\.InProceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue,O\. Pietquin, S\. Muresan, V\. Chen, C\. Kennington, D\. Vandyke, N\. Dethlefs, K\. Inoue, E\. Ekstedt, and S\. Ultes \(Eds\.\),1st virtual meeting,pp\. 53–56\.External Links:[Link](https://aclanthology.org/2020.sigdial-1.7/),[Document](https://dx.doi.org/10.18653/v1/2020.sigdial-1.7)Cited by:[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p1.1)\.
- Kojimaet al\.\(2022\)T\. Kojima, S\. S\. Gu, M\. Reid, Y\. Matsuo, and Y\. IwasawaLarge language models are zero\-shot reasoners\.Advances in neural information processing systems35,pp\. 22199–22213\.Cited by:[§D\.2](https://arxiv.org/html/2608.28884#A4.SS2.p1.1),[§5\.2](https://arxiv.org/html/2608.28884#S5.SS2.SSS0.Px2.p1.1)\.
- Krantiet al\.\(2024\)C\. Kranti, S\. Hakimov, and D\. SchlangenRetrieval\-augmented code generation for situated action generation: a case study on Minecraft\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 11159–11170\.External Links:[Link](https://aclanthology.org/2024.findings-emnlp.652/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-emnlp.652)Cited by:[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p1.1)\.
- Moravec \(1988\)H\. MoravecMind children: the future of robot and human intelligence\.Harvard University Press\.External Links:ISBN 9780674576186,LCCN lc88021343,[Link](https://books.google.com/books?id=56mb7XuSx3QC)Cited by:[§1](https://arxiv.org/html/2608.28884#S1.p1.1)\.
- Narayan\-Chenet al\.\(2019\)A\. Narayan\-Chen, P\. Jayannavar, and J\. HockenmaierCollaborative dialogue in Minecraft\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,A\. Korhonen, D\. Traum, and L\. Màrquez \(Eds\.\),Florence, Italy,pp\. 5405–5415\.External Links:[Link](https://aclanthology.org/P19-1537/),[Document](https://dx.doi.org/10.18653/v1/P19-1537)Cited by:[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p1.1)\.
- Padmakumaret al\.\(2022\)A\. Padmakumar, J\. Thomason, A\. Shrivastava, P\. Lange, A\. Narayan\-Chen, S\. Gella, R\. Piramuthu, G\. Tur, and D\. Hakkani\-TurTEACh: task\-driven embodied agents that chat\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.36,pp\. 2017–2025\.Cited by:[§2\.1](https://arxiv.org/html/2608.28884#S2.SS1.p1.1)\.
- \[13\]MineflayerExternal Links:[Link](https://github.com/PrismarineJS/mineflayer)Cited by:[§E\.2](https://arxiv.org/html/2608.28884#A5.SS2.p4.1),[§5\.1](https://arxiv.org/html/2608.28884#S5.SS1.p1.1)\.
- Shridharet al\.\(2020\)M\. Shridhar, J\. Thomason, D\. Gordon, Y\. Bisk, W\. Han, R\. Mottaghi, L\. Zettlemoyer, and D\. FoxALFRED: a benchmark for interpreting grounded instructions for everyday tasks\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 10740–10749\.Cited by:[§2\.1](https://arxiv.org/html/2608.28884#S2.SS1.p1.1)\.
- Tao \(2024\)T\. TaoMachine assisted proof\.Notices of the American Mathematical Society, to appear\.Cited by:[§1](https://arxiv.org/html/2608.28884#S1.p1.1)\.
- Thomasonet al\.\(2020\)J\. Thomason, M\. Murray, M\. Cakmak, and L\. ZettlemoyerVision\-and\-dialog navigation\.InConference on Robot Learning,pp\. 394–406\.Cited by:[§2\.1](https://arxiv.org/html/2608.28884#S2.SS1.p1.1)\.
- Wanget al\.\(2024\)G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. AnandkumarVoyager: an open\-ended embodied agent with large language models\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=ehfRiF0R3a)Cited by:[Table 1](https://arxiv.org/html/2608.28884#S2.T1.2.2.1)\.
- Wanget al\.\(2026\)M\. Wang, R\. Lin, K\. Hu, J\. Jiao, N\. Chowdhury, E\. Chang, and T\. PatwardhanFrontierScience: evaluating AI’s ability to perform expert\-level scientific tasks\.External Links:2601\.21165,[Link](https://arxiv.org/abs/2601.21165)Cited by:[§1](https://arxiv.org/html/2608.28884#S1.p1.1)\.
- White\*et al\.\(2025\)I\. White\*, K\. Nottingham\*, A\. Maniar, M\. Robinson, H\. Lillemark, M\. Maheshwari, L\. Qin, and P\. AmmanabroluCollaborating action by action: a multi\-agent LLM framework for embodied reasoning\.arXiv preprint arXiv:2504\.17950\.External Links:[Link](https://arxiv.org/abs/2504.17950)Cited by:[§D\.4](https://arxiv.org/html/2608.28884#A4.SS4.p1.1),[§E\.2](https://arxiv.org/html/2608.28884#A5.SS2.p1.1),[§E\.2](https://arxiv.org/html/2608.28884#A5.SS2.p4.1),[Figure 2](https://arxiv.org/html/2608.28884#S1.F2),[§2\.2](https://arxiv.org/html/2608.28884#S2.SS2.p2.1),[Table 1](https://arxiv.org/html/2608.28884#S2.T1.2.4.1),[§5\.1](https://arxiv.org/html/2608.28884#S5.SS1.p1.1)\.
- Wuet al\.\(2025\)Y\. Wu, L\. Wang, and R\. LiuCEQuest: benchmarking large language models for construction estimation\.External Links:2508\.16081,[Link](https://arxiv.org/abs/2508.16081)Cited by:[§2\.3](https://arxiv.org/html/2608.28884#S2.SS3.p1.1),[Table 1](https://arxiv.org/html/2608.28884#S2.T1.2.5.1)\.
- Yanget al\.\(2025\)R\. Yang, H\. Chen, J\. Zhang, M\. Zhao, C\. Qian, K\. Wang, Q\. Wang, T\. V\. Koripella, M\. Movahedi, M\. Li, H\. Ji, H\. Zhang, and T\. ZhangEmbodiedBench: comprehensive benchmarking multi\-modal large language models for vision\-driven embodied agents\.InForty\-second International Conference on Machine Learning,External Links:[Link](https://openreview.net/forum?id=DgGF2LEBPS)Cited by:[Table 1](https://arxiv.org/html/2608.28884#S2.T1.2.3.1)\.

## Appendix AExpert Annotation Details

MineCEraft tasks were created through a structured expert\-guided process as follows\. First, a construction\-engineering faculty member with a Ph\.D\. in Civil Engineering designed the high\-level task structures, focusing on instructions commonly used in real\-world construction practice and that can be objectively evaluated in the game\-simulation world of Minecraft\. For each task family, the expert concretized the intended construction scenarios, constraints, and the feasible programmatic checks\. We then augmented these templates into 723 task instances by varying dimensions, materials, counts, wording, and planning/revision orders, following the expert’s guidelines\. Each task instance is stored as natural\-language prompt variants paired with executable checks\. Lastly, the task set was reviewed again during human evaluation by the authors and the expert to verify that the automatic scoring logic is consistent with the manual inspection results and reflects the intended requirements\.

![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative_abc_chatgpt.png)![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/claude-arch_bridge.png)![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/arch-gemini.jpg)\(a\) ChatGPT 5\.4 Thinking \(Image\)\(b\) Claude Sonnet 4\.6 Extended \(Image\)\(c\) Gemini 3 Thinking \(Image\)\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\[ Roadway \]\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\-\-\-\-\-\-\-\-\-\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_/ Roadway \\\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\-\-\-\-\-\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜˜\#\#\#\#\#\#\#\#\#\#\#\#\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_ ROADWAY \_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\| \|\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\- STONE STRUCTURE \-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\- OPEN ARCH \-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\#\#\#\#\#\-\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\#\#\#\#\#\#\#\#\#\#\#\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\#\#\#\#\#\#\#\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\#\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_/ \\\| \#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\# \| <\-\- Roadway\| \#\#\#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\#\#\# \|\| \#\#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\#\# \|\| \#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\# \| <\-\- The Arch\| \#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\# \|\| \#\#\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\-\#\#\#\# \| <\-\- Abutments\| \#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\#\# \|\|\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\|\(d\) ChatGPT 5\.4 Thinking\(e\) Claude Sonnet 4\.6 Extended\(f\) Gemini 3 ThinkingFigure 5:Qualitative examples of building arch bridges in 2D\. While LLMs can generate arch bridges when prompted with“Draw an arch bridge as an image\.”, they often produce irregular or structurally incorrect shapes when asked in a text\-based format, such as“Let’s build an arch bridge viewed from the side using stones \(\#\) and empty spaces \(\-\)\.”\. By comparison, Gemini 3 can generate a relatively arch\-like structure even in the text\-based setting\.
## Appendix BToy Example: Arch Bridge Challenge

As a point of departure, we begin with a very simple experiment, the*Arch Bridge Challenge*\(ABC\)\. Consider the following prompt:

> Let’s build an arch bridge viewed from the side using stones \(\#\) and empty spaces \(\-\)\.

Chances are that even a state\-of\-the\-art LLM \(e\.g\., ChatGPT 5\) will generate a non\-arch structure \(or not even a bridge at all\), as illustrated in Figure[5](https://arxiv.org/html/2608.28884#A1.F5)\. In contrast, also consider an LLM output when prompted with:

> Draw an arch bridge as an image\.

In general, with the second prompt, a text\-to\-image model can produce a beautiful, well\-formed bridge\. This contrast suggests that certain forms of knowledge within LLMs are not internally well\-connected; in other words, their ability to generalize across modalities is limited\.

Put differently, although an LLM may respond as if it understands what an arch is, merely changing the output modality from image generation to text generation reveals a failure of generalization\. This example, therefore, can serve to debunk the illusion that LLMs trulyunderstandthe structural principles of an arch bridge\.

## Appendix CEvaluation Algorithms

##### Notation and Terminologies\.

We represent a structure as a block setBB, where each blockb∈Bb\\in Bhas integer coordinates\(x,y,z\)\(x,y,z\)and a material labelb\.materialb\.\\mathrm\{material\}\. When construction order is relevant, we writeB=\(b1,…,bn\)B=\(b\_\{1\},\\dots,b\_\{n\}\)for the ordered block\-placement sequence\. We usematch\(b\.material,m\)\\mathrm\{match\}\(b\.\\mathrm\{material\},m\)to denote material matching under the canonical material taxonomy used in the benchmark\. Some checks operate on predefined subsets of blocks\. These are specified using a region selectorrrthat extracts a subsetR=r⁡\(B\)⊆BR=r\(B\)\\subseteq B\(e\.g\., corners at a given height, boundary walls, or the top layer\)\. For connectivity in 3D we use 6\-neighbor face adjacency\. For planar flood\-fill operations on thexx–zzgrid \(e\.g\., room counting\), we use 8\-neighborhood connectivity\.

Concretely, metrics are evaluated by the following algorithms:

- •Material accuracy: Algorithms[1](https://arxiv.org/html/2608.28884#alg1)–[3](https://arxiv.org/html/2608.28884#alg3)\.
- •Shape accuracy: Algorithms[4](https://arxiv.org/html/2608.28884#alg4)–[8](https://arxiv.org/html/2608.28884#alg8)\.
- •Size accuracy: Algorithms[9](https://arxiv.org/html/2608.28884#alg9)and[10](https://arxiv.org/html/2608.28884#alg10)\.
- •Physical plausibility: Algorithm[11](https://arxiv.org/html/2608.28884#alg11)\.
- •Structural stability: Appendix[G](https://arxiv.org/html/2608.28884#A7)\.
- •Efficiency: Algorithm[12](https://arxiv.org/html/2608.28884#alg12)\.
- •Dependency: Algorithm[13](https://arxiv.org/html/2608.28884#alg13)\.

Algorithm 1Global Material Uniformity1:Block set

BB, target material

mm
2:Trueiff every block uses material

mm
3:if

B=∅B=\\emptysetthen

4:returnFalse

5:endif

6:for all

b∈Bb\\in Bdo

7:if

¬match\(b\.material,m\)\\neg\\mathrm\{match\}\(b\.\\mathrm\{material\},m\)then

8:returnFalse

9:endif

10:endfor

11:returnTrue

Algorithm 2Counting Blocks of a Given Material1:Block set

BB, expected material

mm\(optional\), substring flag

ss
2:Number of blocks whose material matches

mm
3:

c​o​u​n​t←0count\\leftarrow 0
4:foreach block

b∈Bb\\in Bdo

5:if

mmisNonethen

6:

c​o​u​n​t←c​o​u​n​t\+1count\\leftarrow count\+1
7:else

8:let

m​a​tmatbe the material of

bb
9:if

ssisTruethen

10:if

mmis a substring of

m​a​tmatthen

11:

c​o​u​n​t←c​o​u​n​t\+1count\\leftarrow count\+1
12:endif

13:else

14:if

m​a​t=mmat=mthen

15:

c​o​u​n​t←c​o​u​n​t\+1count\\leftarrow count\+1
16:endif

17:endif

18:endif

19:endfor

20:return

c​o​u​n​tcount

Algorithm 3Regional Material Ratio Check1:Block set

BB, region selector

rr, target material

mm, threshold

τ\\tau
2:Trueiff the selected region mostly uses material

mm
3:

R←r⁡\(B\)R\\leftarrow r\(B\)
4:if

R=∅R=\\emptysetthen

5:returnFalse

6:endif

7:

c←\|\{b∈R:match\(b\.material,m\)\}\|c\\leftarrow\|\\\{b\\in R:\\mathrm\{match\}\(b\.\\mathrm\{material\},m\)\\\}\|
8:return

\(c/\|R\|\)≥τ\(c/\|R\|\)\\geq\\tau

Algorithm 4Surface Concavity Check1:Block set

BB, mode

m∈\{Top,Bottom\}m\\in\\\{\\textsc\{Top\},\\textsc\{Bottom\}\\\}
2:Trueiff some axis\-aligned slice is concave

3:if

B=∅B=\\emptysetthen

4:returnTrue

5:endif

6:Compute surface height map

S⁡\(x,z\)S\(x,z\)using

7:

maxy\\max\_\{y\}forTopand

miny\\min\_\{y\}forBottom

8:Construct all 1D slices of

SSby fixing

xxor

zz
9:for allslices

PPdo

10:if

PPsatisfies the chord\-based concavity conditionthen

11:returnTrue

12:endif

13:endfor

14:returnFalse

Algorithm 5Room Counting1:Block set

BB, walking height

yy\(default

y=1y=1\), target room count

kk
2:Trueiff at least

kkenclosed rooms exist

3:Let

OObe the set of occupied cells

\(x,z\)\(x,z\)on the

xx–

zzplane at height

yy
4:if

O=∅O=\\emptysetthen

5:returnFalse

6:endif

7:Compute the bounding box of

OOand expand it by one cell in all directions

8:Initialize a queue with all boundary cells of the expanded box that are not in

OO
9:Let

AAbe the set of cells reachable from these boundary cells⊳\\trianglerightOutside air detection using the Flood\-fill algorithm

10:whilethe queue is not emptydo

11:Pop a cell

\(x,z\)\(x,z\)
12:foreach of its 8\-neighbor cells

\(x′,z′\)\(x^\{\\prime\},z^\{\\prime\}\)do

13:if

\(x′,z′\)\(x^\{\\prime\},z^\{\\prime\}\)is inside the bounding box and not in

OOand not in

AAthen

14:Add

\(x′,z′\)\(x^\{\\prime\},z^\{\\prime\}\)to

AAand push it into the queue

15:endif

16:endfor

17:endwhile

18:Count connected components of empty cells that are neither in

OOnor in

AA
19:ifthe number of such enclosed components

≥k\\geq kthen

20:returnTrue

21:else

22:returnFalse

23:endif

Algorithm 6Minimum Inter\-Cluster Span Check1:Block set

BB, height

yy, minimum span threshold

smins\_\{\\min\}, neighborhood optionuse\_8\_neighbors

2:Trueiff every pair of clusters at height

yyis at least

smins\_\{\\min\}apart

3:Extract all occupied cells on the

xx–

zzplane at height

yy
4:Compute connected components

C1,C2,…,CnC\_\{1\},C\_\{2\},\\dots,C\_\{n\}on this plane

5:using 4\-neighborhood by default, or 8\-neighborhood ifuse\_8\_neighborsis enabled

6:if

n<2n<2then

7:returnFalse

8:endif

9:foreach unordered pair of distinct clusters

\(Ci,Cj\)\(C\_\{i\},C\_\{j\}\)do

10:Compute the minimum pairwise Euclidean distance

d⁡\(Ci,Cj\)=minp∈Ciq∈Cj⁡∥p−q∥2d\(C\_\{i\},C\_\{j\}\)=\\min\_\{\\begin\{subarray\}\{c\}p\\in C\_\{i\}\\\\ q\\in C\_\{j\}\\end\{subarray\}\}\\lVert p\-q\\rVert\_\{2\}
11:if

d⁡\(Ci,Cj\)<smind\(C\_\{i\},C\_\{j\}\)<s\_\{\\min\}then

12:returnFalse

13:endif

14:endfor

15:returnTrue

Algorithm 7Flat Exit Path1:Block set

BB, optional start selector

msm\_\{s\}
2:

11iff a flat path to the outside exists from every start

3:if

B=∅B=\\emptysetthen

4:return

11
5:endif

6:

yf←miny⁡\(B\)y\_\{f\}\\leftarrow\\min\_\{y\}\(B\)⊳\\trianglerightfloor level

7:if

yf\>−1y\_\{f\}\>\-1then

8:return

00
9:endif

10:

F←F\\leftarrowwalkable cells at level

yfy\_\{f\}
11:

P←P\\leftarrowstart positions \(matching

msm\_\{s\}, or centroid if none\)

12:for all

p∈Pp\\in Pdo

13:if

y⁡\(p\)≠yf\+1y\(p\)\\neq y\_\{f\}\+1then

14:return

00
15:endif

16:Run BFS on the

xx–

zzplane from

pp
17:allowing moves only within

FF
18:with sufficient head clearance

19:ifno path reaches outside the bounding boxthen

20:return

00
21:endif

22:endfor

23:return

11

Algorithm 8Reachability by Stairs1:Block set

BB, target level

yminy\_\{\\min\}
2:

11iff an agent can reach height

≥ymin\\geq y\_\{\\min\}
3:Initialize BFS from all ground blocks with head clearance

4:whilequeue not emptydo

5:Pop

\(x,y,z\)\(x,y,z\)
6:if

y≥yminy\\geq y\_\{\\min\}then

7:return

11
8:endif

9:Enqueue valid 4\-neighbor moves at level

yyor

y\+1y\+1
10:endwhile

11:return

00

Algorithm 9Bounding\-Box Size Check1:Block set

BB, planar size

\(sx,sz\)\(s\_\{x\},s\_\{z\}\), optional height

sys\_\{y\}, mode

m∈\{Equal,AtMost\}m\\in\\\{\\textsc\{Equal\},\\textsc\{AtMost\}\\\}
2:

11iff the bounding\-box size constraint holds

3:if

B=∅B=\\emptysetthen

4:return

00
5:endif

6:Compute bounding\-box ranges

\(dx,dy,dz\)\(d\_\{x\},d\_\{y\},d\_\{z\}\)
7:

\(rx,rz\)←\(sx−1,sz−1\)\(r\_\{x\},r\_\{z\}\)\\leftarrow\(s\_\{x\}\-1,s\_\{z\}\-1\)
8:if

m=Equalm=\\textsc\{Equal\}then

9:return

11iff

\{dx,dz\}=\{rx,rz\}\\\{d\_\{x\},d\_\{z\}\\\}=\\\{r\_\{x\},r\_\{z\}\\\}and \(if given\)

dy=sy−1d\_\{y\}=s\_\{y\}\-1
10:else

11:return

11iff

\{dx,dz\}≤\{rx,rz\}\\\{d\_\{x\},d\_\{z\}\\\}\\leq\\\{r\_\{x\},r\_\{z\}\\\}and \(if given\)

dy≤sy−1d\_\{y\}\\leq s\_\{y\}\-1
12:endif

Algorithm 10Vertical and Minimum\-Size Check1:Block set

BB, constraint

cc
2:

11iff the size constraint holds

3:if

B=∅B=\\emptysetthen

4:return

00
5:endif

6:Compute

miny⁡\(B\)\\min\_\{y\}\(B\),

maxy⁡\(B\)\\max\_\{y\}\(B\), and

\(dx,dy,dz\)\(d\_\{x\},d\_\{y\},d\_\{z\}\)
7:if

ccspecifies exact minimum depththen

8:return

11iff

miny⁡\(B\)=y⋆\\min\_\{y\}\(B\)=y^\{\\star\}
9:elseif

ccspecifies minimum depththen

10:return

11iff

miny⁡\(B\)≤y⋆\\min\_\{y\}\(B\)\\leq y^\{\\star\}
11:elseif

ccspecifies minimum heightthen

12:return

11iff

maxy⁡\(B\)≥y⋆\\max\_\{y\}\(B\)\\geq y^\{\\star\}
13:elseif

ccspecifies minimum sizethen

14:return

11iff

dx,dy,dz≥s⋆−1d\_\{x\},d\_\{y\},d\_\{z\}\\geq s^\{\\star\}\-1
15:endif

16:return

00

Algorithm 11Ground Connectivity Check1:Block set

BB
2:

11iff every block connects to the ground

3:if

B=∅B=\\emptysetthen

4:return

11
5:endif

6:

G←\{b∈B:y⁡\(b\)≤0\}G\\leftarrow\\\{b\\in B:y\(b\)\\leq 0\\\}
7:if

G=∅G=\\emptysetthen

8:return

00
9:endif

10:Run BFS from all blocks in

GGusing 6\-neighbor adjacency

11:ifall blocks in

BBare visitedthen

12:return

11
13:else

14:return

00
15:endif

Algorithm 12Path Efficiency1:Ordered block sequence

B=\(b1,…,bn\)B=\(b\_\{1\},\\dots,b\_\{n\}\)
2:Efficiency score in

\[0,1\]\[0,1\]
3:if

n≤1n\\leq 1then

4:return

11
5:endif

6:

L←∑i=2n‖bi−bi−1‖1L\\leftarrow\\sum\_\{i=2\}^\{n\}\\\|b\_\{i\}\-b\_\{i\-1\}\\\|\_\{1\}
7:

L⋆←n−1L^\{\\star\}\\leftarrow n\-1
8:if

L≤0L\\leq 0then

9:return

11
10:endif

11:return

min⁡\(1,L⋆/L\)\\min\(1,L^\{\\star\}/L\)

Algorithm 13Material Sequence Order Check1:Ordered block sequence

B=\(b1,…,bn\)B=\(b\_\{1\},\\dots,b\_\{n\}\), target sequence

M=\(m1,…,mk\)M=\(m\_\{1\},\\dots,m\_\{k\}\)
2:

11iff

MMappears in order in

BB
3:if

M=∅M=\\emptysetthen

4:return

11
5:endif

6:if

B=∅B=\\emptysetthen

7:return

00
8:endif

9:

j←1j\\leftarrow 1
10:for

i=1i=1to

nndo

11:if

match\(bi\.material,mj\)\\mathrm\{match\}\(b\_\{i\}\.\\mathrm\{material\},m\_\{j\}\)then

12:

j←j\+1j\\leftarrow j\+1
13:if

j\>kj\>kthen

14:return

11
15:endif

16:endif

17:endfor

18:return

00

## Appendix DAdditional Results and Statistics

### D\.1Overall Problem Accuracy

One may ask the following two questions\. First, at the level of the entire criteria of each problem, how often does a model satisfy all binary accuracy\-related requirements? Second, among the constructions that do satisfy those binary requirements, how high are the graded/continuous metrics, such as efficiency and structural stability?

These questions are particularly important in construction, where even a single mistake is often not allowed\. Since simply averaging heterogeneous metrics can obscure risk, a stricter and more reliable aggregation at the problem level is reported in Table[4](https://arxiv.org/html/2608.28884#A4.T4)\. We first define a construction asaccurateif it satisfies the intersection ofallaccuracy criteria\. We then further aggregate across safety and planning by additionally requiring physical plausibility and dependency satisfaction, which are also binary, producing an overall success criterion that reflects the intersection of these dimensions\. As shown in Table[4](https://arxiv.org/html/2608.28884#A4.T4), the resulting problem\-level success rates are substantially lower than the instruction\-level accuracies reported earlier\. This stricter aggregation makes it clearer that current LLM\-based agents still struggle to perform construction tasks in a reliable manner\.

A second question is whether constructions that areaccuratein this strict binary sense tend to achieve better performance on graded metrics\. Table[5](https://arxiv.org/html/2608.28884#A4.T5)suggests two main observations \(excludingllama\-4\-17b\-16e, which produced no fully accurate constructions on the relevant subsets\)\. First, binary accuracy does not necessarily translate into improved efficiency\. That is, even when a model produces a construction that correctly satisfies the required material, shape, and size constraints, the resulting building process is not consistently more movement\-efficient\.

Second, constructions that satisfy the accuracy criteria sometimes exhibit slightly higher structural stability for some models \(gpt5\-miniandgemini\-3\-pro\)\. This observation is consistent with basic engineering intuition\. There is no inherent reason to expect that a structure built exactly according to specification must also have been constructed efficiently\. In contrast, constructions that satisfy material, shape, and size constraints tend to exhibit slightly higher structural stability\.

llama\-4\-17b\-16eclaude\-4\-5\-sonnetgpt5\-minigemini\-3\-proMetric\#Problems\#CorrectAccuracy\#CorrectAccuracy\#CorrectAccuracy\#CorrectAccuracyOverall Accuracy \(acc∩pp∩dep\\mathrm\{acc\}\\cap\\mathrm\{pp\}\\cap\\mathrm\{dep\}\)72330\.4%19226\.6%28639\.6%27738\.3%Accuracy \(material∩shape∩size\\mathrm\{material\}\\cap\\mathrm\{shape\}\\cap\\mathrm\{size\}\)699111\.6%19728\.2%31745\.4%26638\.1%Material45512928\.4%31469\.0%41090\.1%29665\.1%Shape563112\.0%11520\.4%21337\.8%17831\.6%Size36971\.9%14539\.3%28477\.0%19552\.8%Physical Plausibility72337251\.5%38553\.3%49067\.8%66391\.7%Dependency15685\.1%10466\.7%15398\.1%12278\.2%

Table 4:Problem\-level aggregation of binary success criteria\.A construction is counted asaccurateonly if it satisfies all applicable accuracy requirements for that problem simultaneously, i\.e\., material∩\\capshape∩\\capsize\. The stricter overall criterion additionally requires physical plausibility and dependency satisfaction, i\.e\., accuracy∩\\capphysical plausibility∩\\capdependency\. This evaluation is stricter than the instruction\-level results in Table[2](https://arxiv.org/html/2608.28884#S4.T2): a single problem may contain multiple instructions \(for multi\-turn tasks\) and multiple checks within the same category \(e\.g\., several shape constraints\), so success is counted only when all required conditions within the problem are satisfied together\. The performance drop compared to Table[2](https://arxiv.org/html/2608.28884#S4.T2)more clearly reveals that current LLMs still struggle to reliably execute construction tasks\.llama\-4\-17b\-16eclaude\-4\-5\-sonnetgpt5\-minigemini\-3\-proMetric\#AllAll\#Acc\.Acc\.\-onlyAll\#Acc\.Acc\.\-onlyAll\#Acc\.Acc\.\-onlyAll\#Acc\.Acc\.\-onlyEff\. \(L1\)11247\.1%0–56\.2%9454\.9%55\.6%11055\.1%59\.7%10054\.8%Stab\. \(stress\)2125\.0%0–57\.1%554\.1%50\.7%759\.7%42\.7%1080\.7%

Table 5:Graded metrics evaluated at the problem level\.‘\#All’ denotes the number of applicable problems for each metric and is shared across models\. For each model, ‘All’ reports the mean score across all applicable problems, ‘\#Acc\.’ is the number of problems whose binary accuracy constraints are fully satisfied, and ‘Acc\.\-only’ reports the mean score restricted to that accurate subset\. Accuracy does not consistently improve efficiency, but accurate constructions sometimes exhibit modest gains in structural stability\.
### D\.2Effect of Chain\-of\-Thought Prompting

We further analyze the effect of chain\-of\-thought style prompting on construction performance\. While prompts such as “Let’s think step by step” have been shown to improve reasoning performance in domains such as mathematics\([Kojima et al\., 2022](https://arxiv.org/html/2608.28884#bib.bib4)\), we observe little to no measurable improvement in embodied construction tasks\. In contrast, prompts that explicitly encourage reasoning about construction order or movement planning lead to modest improvements in planning\-related metrics\. Detailed results are reported in Table[6](https://arxiv.org/html/2608.28884#A4.T6)\.

Prompt Typellama \-4 \-scoutclaude \-4\-5 \-sonnetgpt\-5 \-minigemini \-3\-proBaseline54\.8%58\.3%58\.3%58\.3%Step\-by\-step52\.4%58\.3%58\.3%58\.3%Δ\\Delta\(vs\. Baseline\)\-2\.4%0\.0%0\.0%0\.0%Order\-aware52\.4%79\.2%79\.2%100\.0%Δ\\Delta\(vs\. Baseline\)\-2\.4%\+20\.9%\+20\.9%\+41\.7%Table 6:Construction efficiency \(%, higher is better\) for the foundation\-building task under different prompting strategies\. The baseline prompt contains only the task instruction\.Step\-by\-stepadds the generic reasoning cue“Let’s think step by step\.”, whereasOrder\-awareadds the task\-specific cue“Let’s think about the order in which we should place the blocks to optimize the movement path\.”Rows labeledΔ\\Deltashow the change in efficiency relative to the baseline\. The generic reasoning cue does not improve efficiency, while the order\-aware cue yields substantial gains for several models\.
### D\.3Sensitivity to Instruction Verbs

We examine how variations in instruction wording affect model performance\. Specifically, we compare prompts that differ only in the verb used \(e\.g\., “build”, “create”, or “construct”\)\. Despite conveying similar semantic intent, we observe measurable performance differences across several models and evaluation criteria, as in Table[7](https://arxiv.org/html/2608.28884#A4.T7)\. This suggests that LLM behavior on construction tasks can be sensitive to subtle variations in natural\-language instructions\.

Professional abb\.Plain wordsMaterial66\.750\.0Shape66\.766\.7Physical Plausibility100\.0100\.0Structural Stability66\.716\.7Dependency66\.7100\.0
“build”“construct”“create”Material50\.0100\.025\.0Shape50\.0100\.050\.0Dependency75\.0100\.075\.0Physical Plausibility100\.0100\.0100\.0Structural Stability25\.050\.050\.0

Table 7:Sensitivity to instruction wording forclaude\-4\-5\-sonnetevaluated on 72 construction tasks of roofs and columns in Section[3\.2\.1](https://arxiv.org/html/2608.28884#S3.SS2.SSS1)\. Top: comparison between a field\-specific abbreviation \(SOMD\) and its plain\-language equivalent \(“slab on metal deck”\) shows little difference in performance, suggesting that construction\-specific terminology is reasonably well captured\. Bottom: prompts differing only in the verb \(“build”, “construct”, “create”\) show measurable differences\.
### D\.4Sensitivity to System Prompt

We further examine whether the system prompt affects model performance\. In particular, we remove the instruction “Be very brief in your responses” from the original MINDcraft\-based system prompt\([White\* et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib16)\)while keeping the remaining setup unchanged\. Table[8](https://arxiv.org/html/2608.28884#A4.T8)compares the original prompt with this modified prompt for four representative models\.

Removing the brevity constraint substantially improves dependency planning across all four models, with gains ranging from 20\.0 to 40\.0 percentage points\. However, the effect is not uniformly positive across other criteria\. For example, some models improve substantially in structural stability, whereas shape, physical plausibility, or efficiency can decrease\. These results indicate that construction performance can be sensitive to system\-prompt design, motivating a more systematic study of prompt design and optimization in future work\.

ModelSystem PromptMaterialShapeSizeP\.P\.S\.S\.Eff\.Dep\.llama\-4\-scout\-17b\-16eOriginal20\.623\.824\.685\.716\.737\.80\.0w/o brevity constraint33\.329\.516\.783\.30\.084\.220\.0Δ\\Delta\+12\.7\+5\.7\-7\.9\-2\.4\-16\.7\+46\.4\+20\.0claude\-4\-5\-sonnetOriginal84\.160\.583\.383\.355\.654\.960\.0w/o brevity constraint90\.547\.591\.766\.791\.953\.1100\.0Δ\\Delta\+6\.4\-13\.0\+8\.4\-16\.6\+36\.3\-1\.8\+40\.0gpt\-5\-miniOriginal87\.152\.583\.359\.267\.154\.962\.2w/o brevity constraint95\.245\.975\.060\.072\.453\.1100\.0Δ\\Delta\+8\.1\-6\.6\-8\.3\+0\.8\+5\.3\-1\.8\+37\.8gemini\-3\-proOriginal71\.447\.538\.097\.849\.469\.860\.0w/o brevity constraint81\.065\.650\.080\.079\.453\.1100\.0Δ\\Delta\+9\.6\+18\.1\+12\.0\-17\.8\+30\.0\-16\.7\+40\.0

Table 8:Sensitivity to the brevity constraint in the system prompt\. We compare the original MINDcraft\-based system prompt with a variant that removes the instruction “Be very brief in your responses\.” Values are percentages, andΔ\\Deltadenotes the change relative to the original prompt\.
### D\.5Additional Multimodal Model Comparison

To further investigate the potential of multimodal models, we additionally compare GPT\-4 with GPT\-4o\. As shown in Table[9](https://arxiv.org/html/2608.28884#A4.T9), GPT\-4o outperforms GPT\-4 across all evaluation criteria, with particularly large improvements in material accuracy and structural stability\. Although this comparison does not directly evaluate visual perception, as visual observations are not provided during evaluation, the result provides indirect evidence that capabilities of multimodal models may still benefit embodied construction in a text\-based interaction setting\.

ModelMaterialShapeSizeP\.P\.S\.S\.Eff\.Dep\.GPT\-442\.9%50\.8%50\.0%70\.0%0\.0%53\.1%40\.0%GPT\-4o85\.7%55\.9%61\.1%84\.4%67\.8%62\.0%60\.0%

Table 9:Additional GPT\-4 vs\. GPT\-4o comparison\.

## Appendix EExperimental Details

### E\.1Human Evaluation

To ensure a fair comparison between humans and LLMs, we carefully controlled the information provided to participants\. Human subjects were given only the following guideline:

Youarebuilder,aconstructionengineerandalsoaworkerataconstructionsite\.WeareintheworldofMinecraft,butyoushouldassumethat,basedonyourcoordinatesandbuildorder,thestructuremaybebuiltinreallifevia3Dprinting\.

This instruction is identical to the system prompt provided to the LLM, except for the skill\-related descriptions\. Since human participants were already familiar with Minecraft controls and had sufficient building materials in their inventories at the start of each task, explanations of the skill library were unnecessary for them\.

We conducted experiments with two voluntary participants using MineCEraft\-Lite\. Both participants are professional construction engineers with three years of field experience in the United States, as well as prior experience playing Minecraft\. The evaluation took an average of 3 hours per participant to complete\. Unlike LLM evaluation, it is impossible to reset context \(or memory of prior tasks\) for human subjects\. To alleviate this issue, we randomized the task order and used different random seeds for each participant\.

### E\.2LLM Agent Details

The LLM was initialized with the following system prompt, which is adapted from the originalmindcraft prompt\([White\* et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib16)\), with minimal modifications to specify construction tasks:

Youarebuilder,aconstructionengineerandalsoaworkerataconstructionsite\.WeareintheworldofMinecraft,butyoushouldassumethat,basedonyourcoordinatesandbuildorder,thestructuremaybebuiltinreallifevia3Dprinting\.UseexpertiseasahighlyskilledcivilandconstructionengineerandcommandsinCOMMANDDOCS\.Beverybriefinyourresponses,don’tapologizeconstantly,don’tgiveinstructionsormakelistsunlessasked,anddon’trefuserequests\.DoNOTaskanyquestions\(e\.g\.,aboutdimensions,location,orwhethertobuild\)\.Startbuildingimmediatelyfromyourcurrentposition\.Don’tpretendtoact;usecommandsimmediatelywhenrequested\.DoNOTsaythis:’Sure,I’vestopped\.’,insteadsaythis:’Sure,I’llstop\.\!stop’\.DoNOTsaythis:’Onmyway\!Givemeamoment\.’,insteadsaythis:’Onmyway\!\!goToPlayer\("playername",3\)’\.Whenconstructionisfinished,say"I’vesuccessfullybuilt\."Respondonlyasbuilder,neveroutput’\(FROMOTHERBOT\)’orpretendtobesomeoneelse\.Ifyouhavenothingtosayordo,respondwithanjustatab’’\.Thisisextremelyimportanttome,takeadeepbreathandhavefun:\)

Additionally, the LLM was given access to a skill library,`COMMAND DOCS`, which specifies executable actions and their arguments:

\*COMMANDDOCS

Youcanusethefollowingcommandstoperformactionsandgetinformationabouttheworld\.

Usethecommandswiththesyntax:\!commandNameor\!commandName\("arg1",1\.2,\.\.\.\)ifthecommandtakesarguments\.

Donotusecodeblocks\.Usedoublequotesforstrings\.Onlyuseonecommandineachresponse,trailingcommandsandcommentswillbeignored\.

\!stats:Getyourbot’slocation,health,hunger,andtimeofday\.

\!inventory:Getyourbot’sinventory\.

\!nearbyBlocks:Gettheblocksnearthebot\.

\!craftable:Getthecraftableitemswiththebot’sinventory\.

\!entities:Getthenearbyplayersandentities\.

\!modes:Getallavailablemodesandtheirdocsandseewhichareon/off\.

\!savedPlaces:Listallsavedlocations\.

\!getCraftingPlan:Providesacomprehensivecraftingplanforaspecifieditem\.Thisincludesabreakdownofrequiredingredients,theexactquantitiesneeded,andananalysisofmissingingredientsorextraitemsneededbasedonthebot’scurrentinventory\.

Params:

targetItem:\(string\)Theitemthatwearetryingtocraft

quantity:\(number\)Thequantityoftheitemthatwearetryingtocraft

\!searchWiki:SearchtheMinecraftWikiforthegivenquery\.

Params:

query:\(string\)Thequerytosearchfor\.

\!help:Listsallavailablecommandsandtheirdescriptions\.

\!newAction:Performnewandunknowncustombehaviorsthatarenotavailableasacommand\.

Params:

prompt:\(string\)Anaturallanguageprompttoguidecodegeneration\.Makeadetailedstep\-by\-stepplan\.

\!stop:Forcestopallactionsandcommandsthatarecurrentlyexecuting\.

\!stfu:Stopallchattingandselfprompting,butcontinuecurrentaction\.

\!restart:Restarttheagentprocess\.

\!clearChat:Clearthechathistory\.

\!goToPlayer:Gotothegivenplayer\.

Params:

player\_name:\(string\)Thenameoftheplayertogoto\.

closeness:\(number\)Howclosetogettotheplayer\.

\!followPlayer:Endlesslyfollowthegivenplayer\.

Params:

player\_name:\(string\)nameoftheplayertofollow\.

follow\_dist:\(number\)Thedistancetofollowfrom\.

\!goToCoordinates:Gotothegivenx,y,zlocation\.

Params:

x:\(number\)Thexcoordinate\.

y:\(number\)Theycoordinate\.

z:\(number\)Thezcoordinate\.

closeness:\(number\)Howclosetogettothelocation\.

\!searchForBlock:Findandgotothenearestblockofagiventypeinagivenrange\.

Params:

type:\(string\)Theblocktypetogoto\.

search\_range:\(number\)Therangetosearchfortheblock\.Minimum32\.

\!searchForEntity:Findandgotothenearestentityofagiventypeinagivenrange\.

Params:

type:\(string\)Thetypeofentitytogoto\.

search\_range:\(number\)Therangetosearchfortheentity\.

\!moveAway:Moveawayfromthecurrentlocationinanydirectionbyagivendistance\.

Params:

distance:\(number\)Thedistancetomoveaway\.

\!rememberHere:Savethecurrentlocationwithagivenname\.

Params:

name:\(string\)Thenametorememberthelocationas\.

\!goToRememberedPlace:Gotoasavedlocation\.

Params:

name:\(string\)Thenameofthelocationtogoto\.

\!givePlayer:Givethespecifieditemtothegivenplayer\.

Params:

player\_name:\(string\)Thenameoftheplayertogivetheitemto\.

item\_name:\(string\)Thenameoftheitemtogive\.

num:\(number\)Thenumberofitemstogive\.

\!consume:Eat/drinkthegivenitem\.

Params:

item\_name:\(string\)Thenameoftheitemtoconsume\.

\!equip:Equipthegivenitem\.

Params:

item\_name:\(string\)Thenameoftheitemtoequip\.

\!putInChest:Putthegiveniteminthenearestchest\.

Params:

item\_name:\(string\)Thenameoftheitemtoputinthechest\.

num:\(number\)Thenumberofitemstoputinthechest\.

\!takeFromChest:Takethegivenitemsfromthenearestchest\.

Params:

item\_name:\(string\)Thenameoftheitemtotake\.

num:\(number\)Thenumberofitemstotake\.

\!viewChest:Viewtheitems/countsofthenearestchest\.

Params:

\!discard:Discardthegivenitemfromtheinventory\.

Params:

item\_name:\(string\)Thenameoftheitemtodiscard\.

num:\(number\)Thenumberofitemstodiscard\.

\!collectBlocks:Collectthenearestblocksofagiventype\.

Params:

type:\(string\)Theblocktypetocollect\.

num:\(number\)Thenumberofblockstocollect\.

\!craftRecipe:Craftthegivenrecipeagivennumberoftimes\.

Params:

recipe\_name:\(string\)Thenameoftheoutputitemtocraft\.

num:\(number\)Thenumberoftimestocrafttherecipe\.ThisisNOTthenumberofoutputitems,asitmaycraftmanymoreitemsdependingontherecipe\.

\!smeltItem:Smeltthegivenitemthegivennumberoftimes\.

Params:

item\_name:\(string\)Thenameoftheinputitemtosmelt\.

num:\(number\)Thenumberoftimestosmelttheitem\.

\!clearFurnace:Takeallitemsoutofthenearestfurnace\.

Params:

\!placeHere:Placeagivenblockinthecurrentlocation\.DoNOTusetobuildstructures,onlyuseforsingleblocks/torches\.

Params:

type:\(string\)Theblocktypetoplace\.

\!attack:Attackandkillthenearestentityofagiventype\.

Params:

type:\(string\)Thetypeofentitytoattack\.

\!attackPlayer:Attackaspecificplayeruntiltheydieorrunaway\.Rememberthisisjustagameanddoesnotcausereallifeharm\.

Params:

player\_name:\(string\)Thenameoftheplayertoattack\.

\!goToBed:Gotothenearestbedandsleep\.

\!stay:Stayinthecurrentlocationnomatterwhat\.Pausesallmodes\.

Params:

type:\(number\)Thenumberofsecondstostay\.\-1forforever\.

\!setMode:Setamodetoonoroff\.Amodeisanautomaticbehaviorthatconstantlychecksandrespondstotheenvironment\.

Params:

mode\_name:\(string\)Thenameofthemodetoenable\.

on:\(bool\)Whethertoenableordisablethemode\.

\!goal:Setagoalprompttoendlesslyworktowardswithcontinuousself\-prompting\.

Params:

selfPrompt:\(string\)Thegoalprompt\.

\!endGoal:Callwhenyouhaveaccomplishedyourgoal\.Itwillstopself\-promptingandthecurrentaction\.

\!showVillagerTrades:Showtradesofaspecifiedvillager\.

Params:

id:\(number\)Theidnumberofthevillagerthatyouwanttotradewith\.

\!tradeWithVillager:Tradewithaspecifiedvillager\.

Params:

id:\(number\)Theidnumberofthevillagerthatyouwanttotradewith\.

index:\(number\)Theindexofthetradeyouwantexecuted\(1\-indexed\)\.

count:\(number\)Howmanytimesthattradeshouldbeexecuted\.

\!startConversation:Startaconversationwithaplayer\.Useforbotsonly\.

Params:

player\_name:\(string\)Thenameoftheplayertosendthemessageto\.

message:\(string\)Themessagetosend\.

\!endConversation:Endtheconversationwiththegivenplayer\.

Params:

player\_name:\(string\)Thenameoftheplayertoendtheconversationwith\.

\!lookAtPlayer:Lookataplayerorlookinthesamedirectionastheplayer\.

Params:

player\_name:\(string\)Nameofthetargetplayer

direction:\(string\)Howtolook\("at":lookattheplayer,"with":lookinthesamedirectionastheplayer\)

\!lookAtPosition:Lookatspecifiedcoordinates\.

Params:

x:\(number\)xcoordinate

y:\(number\)ycoordinate

z:\(number\)zcoordinate

\!digDown:Digsdownaspecifieddistance\.Willstopifitreacheslava,water,orafallof\>=4blocksbelowthebot\.

Params:

distance:\(number\)Distancetodigdown

\!useOn:Use\(rightclick\)thegiventoolonthenearesttargetofthegiventype\.

Params:

tool\_name:\(string\)Nameofthetooltouse,or"hand"fornotool\.

target:\(string\)Thetargetasanentitytype,blocktype,or"nothing"fornotarget\.

For each task, the LLM was provided with a problem description as a user prompt\. After completing each multi\-turn interaction scenario, the conversation history was reset before proceeding to the next task to prevent cross\-task information leakage\. We used each model provider’s default sampling parameters \(e\.g\., temperature\) for all experiments\.

The LLM interacts with the Minecraft world via Mineflayer\([PrismarineJS,](https://arxiv.org/html/2608.28884#bib.bib15)\), which provides low\-level actions such asplaceBlock\(from, to\)\.mindcraft\([White\* et al\., 2025](https://arxiv.org/html/2608.28884#bib.bib16)\)enables high\-level planning by allowing the LLM to decompose prompts into sequences of executable low\-level actions\. In themindcraft builder setup, every optional behavior is turned off during the evaluation: self\-preservation, unstuck, cowardice, self\-defense, hunting, item collecting, torch placing, elbow room, and idle staring\. However,`placeBlock`/`breakBlockAt`is allowed for the bot to place/remove blocks instantly via the`setblock`command in Minecraft, rather than through the`pathfinder`of Mineflayer, enabling faster construction for large\-scale evaluation\.

For evaluation, action logs generated during execution are parsed into ordered lists of spatial coordinates\. These coordinates are then automatically graded according to the logic described in Section[4](https://arxiv.org/html/2608.28884#S4)\.

## Appendix FRepresentativeness of MineCEraft\-Lite

CategorySubcategoryModelsHumanllama\-4 \-scout\-17b\-16eclaude\-4\-5 \-sonnetgpt\-5\-minigemini\-3 \-proAccuracyMaterial20\.6%\(≈27\.0\)±7\.2\{\}\_\{\\pm 7\.2\}^\{\(\\approx 27\.0\)\}84\.1%\(≈84\.5\)±2\.4\{\}\_\{\\pm 2\.4\}^\{\(\\approx 84\.5\)\}87\.1%\(≈92\.7\)±7\.1\{\}\_\{\\pm 7\.1\}^\{\(\\approx 92\.7\)\}71\.4%\(≈70\.4\)±4\.8\{\}\_\{\\pm 4\.8\}^\{\(\\approx 70\.4\)\}97\.8%±2\.1Shape23\.8%\(≈22\.2\)±7\.1\{\}\_\{\\pm 7\.1\}^\{\(\\approx 22\.2\)\}60\.5%\(≈55\.8\)±3\.6\{\}\_\{\\pm 3\.6\}^\{\(\\approx 55\.8\)\}52\.5%\(≈53\.4\)±4\.1\{\}\_\{\\pm 4\.1\}^\{\(\\approx 53\.4\)\}47\.5%\(≈50\.3\)±4\.0\{\}\_\{\\pm 4\.0\}^\{\(\\approx 50\.3\)\}99\.2%±0\.8Size24\.6%\(≈3\.2\)±12\.0\{\}\_\{\\pm 12\.0\}^\{\(\\approx 3\.2\)\}83\.3%\(≈68\.1\)±16\.7\{\}\_\{\\pm 16\.7\}^\{\(\\approx 68\.1\)\}83\.3%\(≈77\.0\)±8\.4\{\}\_\{\\pm 8\.4\}^\{\(\\approx 77\.0\)\}38\.0%\(≈52\.8\)±19\.5\{\}\_\{\\pm 19\.5\}^\{\(\\approx 52\.8\)\}91\.7%±0\.0SafetyPhysical Plausibility85\.7%\(≈77\.8\)±3\.2\{\}\_\{\\pm 3\.2\}^\{\(\\approx 77\.8\)\}83\.3%\(≈85\.0\)±0\.0\{\}\_\{\\pm 0\.0\}^\{\(\\approx 85\.0\)\}59\.2%\(≈68\.3\)±13\.4\{\}\_\{\\pm 13\.4\}^\{\(\\approx 68\.3\)\}97\.8%\(≈92\.0\)±3\.4\{\}\_\{\\pm 3\.4\}^\{\(\\approx 92\.0\)\}98\.3%±1\.7Structural Stability16\.7%\(≈25\.0\)±25\.0\{\}\_\{\\pm 25\.0\}^\{\(\\approx 25\.0\)\}55\.6%\(≈57\.1\)±17\.4\{\}\_\{\\pm 17\.4\}^\{\(\\approx 57\.1\)\}67\.1%\(≈50\.7\)±22\.3\{\}\_\{\\pm 22\.3\}^\{\(\\approx 50\.7\)\}49\.4%\(≈42\.7\)±19\.7\{\}\_\{\\pm 19\.7\}^\{\(\\approx 42\.7\)\}93\.9%±6\.1PlanningEfficiency37\.8%\(≈47\.1\)±17\.8\{\}\_\{\\pm 17\.8\}^\{\(\\approx 47\.1\)\}54\.9%\(≈56\.2\)±1\.2\{\}\_\{\\pm 1\.2\}^\{\(\\approx 56\.2\)\}54\.9%\(≈55\.6\)±1\.2\{\}\_\{\\pm 1\.2\}^\{\(\\approx 55\.6\)\}69\.8%\(≈59\.7\)±12\.9\{\}\_\{\\pm 12\.9\}^\{\(\\approx 59\.7\)\}95\.6%±2\.9Dependency0\.0%\(≈7\.4\)±0\.0\{\}\_\{\\pm 0\.0\}^\{\(\\approx 7\.4\)\}60\.0%\(≈96\.3\)±0\.0\{\}\_\{\\pm 0\.0\}^\{\(\\approx 96\.3\)\}62\.2%\(≈68\.6\)±20\.0\{\}\_\{\\pm 20\.0\}^\{\(\\approx 68\.6\)\}60\.0%\(≈78\.2\)±20\.0\{\}\_\{\\pm 20\.0\}^\{\(\\approx 78\.2\)\}90\.0%±10\.0Table 10:Performance on MineCEraft\-Lite compared with the full benchmark\.\. Reported values are means on Lite; subscripts \(±\) indicate half of the min–max range across three random seeds \(or, for humans, across different participants\)\. Superscripts \(≈\) denote the corresponding mean on the full benchmark dataset\. Although the full benchmark provides more stable estimates due to its larger task set, the values mostly remain within the Lite min–max range\.To assess how well the Lite version represents the full dataset, we evaluated the four models reported in the main results on MineCEraft\-Lite using three different random seeds and compared the results with those from the full dataset\. As in Table[10](https://arxiv.org/html/2608.28884#A6.T10), although the scores vary slightly depending on the random seed, the Lite results generally fall within the range of the full\-dataset results, suggesting that the Lite version provides a reasonable approximation of the full benchmark\. For more experimental results on MineCEraft\-Lite, please refer to Table[11](https://arxiv.org/html/2608.28884#A6.T11)\.

ModelAccuracySafetyPlanningMaterialShapeSizePhysical PlausibilityStructural StabilityEfficiencyDependencyQwen3\-32B34\.9%±2\.445\.2%±4\.141\.7%±16\.770\.0%±3\.318\.3%±13\.454\.2%±1\.420\.0%±20\.0llama\-4\-scout\-17b\-16e\-Instruct20\.6%±7\.223\.8%±7\.124\.6%±12\.085\.7%±3\.216\.7%±25\.037\.8%±17\.80\.0%±0\.0llama\-3\.3\-70b\-Instruct66\.7%±7\.254\.2%±5\.977\.8%±12\.585\.6%±5\.063\.5%±23\.454\.1%±1\.360\.0%±0\.0claude\-haiku\-4\-5\-2025100179\.4%±4\.859\.6%±10\.775\.0%±8\.385\.6%±5\.065\.7%±6\.654\.2%±1\.460\.0%±0\.0claude\-opus\-4\-5\-2025110195\.2%±0\.064\.0%±5\.088\.9%±4\.280\.0%±10\.067\.6%±17\.454\.2%±1\.473\.3%±10\.0claude\-sonnet\-4\-5\-2025092984\.1%±2\.460\.5%±3\.683\.3%±16\.783\.3%±0\.055\.6%±17\.454\.9%±1\.260\.0%±0\.0claude\-sonnet\-4\-690\.5%±4\.862\.3%±4\.177\.8%±12\.566\.7%±5\.066\.6%±16\.754\.2%±1\.486\.7%±10\.0gpt\-5\-mini87\.1%±7\.152\.5%±4\.183\.3%±8\.459\.2%±13\.467\.1%±22\.354\.9%±1\.262\.2%±20\.0gpt\-5\-nano71\.4%±11\.944\.2%±7\.541\.7%±8\.457\.5%±37\.017\.7%±11\.347\.5%±11\.553\.3%±20\.0gpt\-4o85\.7%±0\.055\.9%±7\.461\.1%±20\.884\.4%±1\.767\.8%±27\.562\.0%±13\.160\.0%±0\.0o392\.0%±4\.854\.8%±7\.477\.8%±4\.273\.3%±3\.455\.4%±18\.954\.2%±1\.460\.0%±0\.0gemini\-2\.5\-flash76\.2%±19\.139\.8%±4\.136\.1%±4\.277\.8%±5\.039\.5%±25\.849\.1%±6\.546\.7%±20\.0gemini\-2\.5\-pro79\.4%±14\.348\.5%±4\.477\.8%±8\.361\.1%±23\.452\.7%±16\.554\.2%±1\.440\.0%±20\.0gemini\-3\-flash\-preview69\.8%±21\.548\.8%±19\.780\.6%±12\.564\.4%±38\.377\.8%±2\.554\.2%±1\.480\.0%±0\.0gemini\-3\-pro\-preview71\.4%±4\.847\.5%±4\.038\.0%±19\.597\.8%±3\.449\.4%±17\.969\.8%±12\.960\.0%±20\.0Table 11:Performance on MineCEraft\-Lite across evaluation categories for multiple LLMs\.
## Appendix GVon Mises Stress Computation

This section summarizes the computation of von Mises stress used to evaluate the structural stability of Minecraft structures in this work\. Our formulation and material parameters follow the elasticity\-based approach of[Beck et al\. \(2024\)](https://arxiv.org/html/2608.28884#bib.bib13)\.222Compared to the original implementation, we cache reference\-configuration quantities, such as edge list, kernel values, and inverse correction matricesA−1A^\{\-1\}, and use NumPy vectorization \(e\.g\., batched tensor products and scatter\-add\); this substantially improves runtime in our evaluation code\.

##### Deformation and Strain

Let𝐗∈ℝ3\\mathbf\{X\}\\in\\mathbb\{R\}^\{3\}denote a material point in the reference configuration and𝐱∈ℝ3\\mathbf\{x\}\\in\\mathbb\{R\}^\{3\}its deformed position\. The deformation gradient𝐅\\mathbf\{F\}is given by

𝐅i​j=∂xi∂Xj,\\mathbf\{F\}\_\{ij\}=\\frac\{\\partial x\_\{i\}\}\{\\partial X\_\{j\}\},\(1\)which locally linearizes the deformation\. From𝐅\\mathbf\{F\}, we compute the Green–Lagrange strain tensor

𝐄=12​\(𝐅⊤​𝐅−𝐈\),\\mathbf\{E\}=\\frac\{1\}\{2\}\\left\(\\mathbf\{F\}^\{\\top\}\\mathbf\{F\}\-\\mathbf\{I\}\\right\),\(2\)which measures changes in squared lengths and is symmetric by construction\.

##### Deformation Gradient Approximation

In Minecraft environment, We use a meshfree Smoothed Particle Hydrodynamics \(SPH\) formulation, in which the structure is discretized into particles without an explicit mesh and deformation is approximated using kernel\-weighted neighbor interactions\.

Specifically, we first compute the approximate deformation gradient estimate

𝐅i⋆=∑j\(𝐱j−𝐱i\)⊗∇W​\(𝐗j−𝐗i\),\\mathbf\{F\}\_\{i\}^\{\\star\}=\\sum\_\{j\}\(\\mathbf\{x\}\_\{j\}\-\\mathbf\{x\}\_\{i\}\)\\otimes\\nabla W\(\\mathbf\{X\}\_\{j\}\-\\mathbf\{X\}\_\{i\}\),\(3\)and apply a first\-order consistency correction

𝐅i\\displaystyle\\mathbf\{F\}\_\{i\}=𝐅i⋆​𝐀i−1,\\displaystyle=\\mathbf\{F\}\_\{i\}^\{\\star\}\\mathbf\{A\}\_\{i\}^\{\-1\},\(4\)𝐀i\\displaystyle\\mathbf\{A\}\_\{i\}=∑j\(𝐗j−𝐗i\)⊗∇W​\(𝐗j−𝐗i\)\.\\displaystyle=\\sum\_\{j\}\(\\mathbf\{X\}\_\{j\}\-\\mathbf\{X\}\_\{i\}\)\\otimes\\nabla W\(\\mathbf\{X\}\_\{j\}\-\\mathbf\{X\}\_\{i\}\)\.
where𝐀i\\mathbf\{A\}\_\{i\}is the reference moment \(correction\) matrix for first\-order consistency\.

##### Stress Measures

Assuming a linear elastic constitutive model, the second Piola–Kirchhoff stress tensor𝐒\\mathbf\{S\}is computed as

𝐒=λ​tr​\(𝐄\)​𝐈\+2​μ​𝐄,\\mathbf\{S\}=\\lambda\\,\\mathrm\{tr\}\(\\mathbf\{E\}\)\\mathbf\{I\}\+2\\mu\\mathbf\{E\},\(5\)whereλ\\lambdaandμ\\muare the Lamé parameters\. The first Piola–Kirchhoff stress𝐏\\mathbf\{P\}is obtained as

𝐏=𝐅𝐒\.\\mathbf\{P\}=\\mathbf\{F\}\\mathbf\{S\}\.\(6\)Finally, the Cauchy stress tensor𝝈\\boldsymbol\{\\sigma\}is given by

𝝈=1J​𝐏𝐅⊤,J=det\(𝐅\)\.\\boldsymbol\{\\sigma\}=\\frac\{1\}\{J\}\\mathbf\{P\}\\mathbf\{F\}^\{\\top\},\\qquad J=\\det\(\\mathbf\{F\}\)\.\(7\)
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/vm_example.png)Figure 6:Example of von Mises stress distribution in a bridge\-like structure\.The peak stress appears near the folded \(bending\) region of the arch rather than around the ground supports where the dead load is concentrated, to evaluate the stress associated with deformation and structural failure\.
##### Von Mises Stress

To obtain a scalar measure of stress magnitude, we compute the deviatoric stress

𝐬=𝝈−13​tr​\(𝝈\)​𝐈,\\mathbf\{s\}=\\boldsymbol\{\\sigma\}\-\\frac\{1\}\{3\}\\mathrm\{tr\}\(\\boldsymbol\{\\sigma\}\)\\mathbf\{I\},\(8\)and define the von Mises stress as

σvm=32​∥𝐬∥F\.\\sigma\_\{\\mathrm\{vm\}\}=\\sqrt\{\\frac\{3\}\{2\}\}\\,\\lVert\\mathbf\{s\}\\rVert\_\{F\}\.\(9\)The von Mises stress is used throughout the paper as a quantitative indicator of structural safety and stress\-based stability assessment, serving as an effective measure of potential failure risk; an illustrative example is shown in Figure[6](https://arxiv.org/html/2608.28884#A7.F6)\.

## Appendix HQualitative Analysis

![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative/frame.png)\(a\)FrameFigure 7:Qualitative Analysis of LLM vs Human Behavior![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative/intent.png)\(a\)Intent
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative/creativity.png)\(b\)Accessibility

Figure 8:Qualitative Analysis of LLM vs Human Behavior \(cont\.\)![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative/functionality.png)\(a\)Functionality
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative/accessability.png)\(b\)Accessibility

Figure 9:Qualitative Analysis of LLM vs Human Behavior \(cont\.\)![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative/res_opt.png)\(a\)Accessibility
![Refer to caption](https://arxiv.org/html/2608.28884v1/imgs/qualitative/dome.png)\(b\)Functionality

Figure 10:Qualitative Analysis of LLM vs Human Behavior \(cont\.\)In this section, we present qualitative examples of the tasks considered in this paper\. Figure[7](https://arxiv.org/html/2608.28884#A8.F7)shows representative cases from the MineCEraft benchmark, highlighting behavioral differences between human participants and LLM\-based agents\.