acceptodds
Under review as a conference paper at ICLR 2027

Factorio Design Benchmark: Evaluating LLM Agents on Hierarchical Design Synthesis

Abstract

Designing a circuit to meet functional requirements or a factory to meet production targets exemplifies hierarchical design synthesis. These tasks require selecting component types and quantities, configuring and placing instances, organizing them into subsystems, and coordinating interfaces and shared infrastructure under coupled feasibility constraints and system-level objectives. Existing benchmarks evaluate important aspects of engineering design, but provide limited evidence on how well large language model (LLM) agents jointly coordinate these decisions across a complete system. We introduce Factorio Design Benchmark, which is based on Factorio, a deterministic factory-construction game. To complete the task, agents are required to submit complete declarative JSON factory designs for static validation, game execution, and production monitoring to decide whether it reaches the target production rate. We evaluate four LLMs on 32 tasks (8 target products 2 product-specific target production rates 2 map settings), with three independent runs per model–task pair. Across all runs, 73.7% produce the target product, but only 40.9% reach the target rate. Beyond these aggregate results, we examine how success varies with task demands. We derive a five-dimensional difficulty profile from task specifications, covering recipe complexity, capacity, logistics, spatial layout, and power infrastructure. The average of its five scores correlates negatively with the proportion of successful runs per task for each model, with Spearman when runs are pooled across all four models. This association supports the profile as a proxy for relative task difficulty within the suite. To identify recurring design failures, we also analyze validation feedback. Entity overlap and unpowered consumers have the highest reported violation counts, with larger counts at high target production rates. These failures highlight challenges in coordinating spatial layouts and shared infrastructure across the complete factory.

Primary area:
datasets and benchmarks
Keywords:
LLM agents, benchmarking, hierarchical design synthesis, factory design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_oY6yOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Factorio Design Benchmark, a benchmark for evaluating LLM agents on hierarchical design synthesis. The benchmark contains 32 factory-design tasks requiring agents to generate complete declarative JSON designs under production, spatial, logistics, and power constraints. Evaluation combines deterministic static validation with execution in Factorio to measure feasibility, production throughput, and design cost. Experiments on four LLMs demonstrate substantial challenges in meeting target production rates. The authors further introduce a five-dimensional task-difficulty profile and analyze common design failures and cost efficiency.

Critical strengths

Firstly, from my sides, this paper is interesting. It doesnt just use agent/llm to design a single task/product/satifsfy some constraints.
Here are the strengths

  1. The paper addresses an practically motivated problem involving system-level design under coupled constraints. The problem is very important in transportation engineering/ OR.
  2. The benchmark provides an objective and executable evaluation framework combining static constraint checking with simulated production throughput.
  3. The experiments include repeated runs, task-difficulty analysis, failure diagnostics, and cost comparisons, offering useful insights into current agents' limitations.
  4. The task specifications, validation rules, and implementation details support reproducibility.
Critical weaknesses
  1. Insufficient Evidence of Task Feasibility. The paper reports that 12 out of 32 tasks remain unsolved, but does not sufficiently establish that all tasks admit feasible solutions satisfying their production and spatial constraints. Without independently verified reference solutions or feasibility evidence, it is difficult to determine whether failures on the hardest tasks reflect agents' limitations or potentially infeasible task specifications. This directly affects the validity of the reported success rates and difficulty analysis.

  2. Missing Planning and Optimization Baselines.The benchmark involves structured production planning, equipment allocation, spatial layout, and logistics optimization, yet the evaluation only considers LLM agents. Comparisons with heuristic, constraint-programming, or hybrid optimization approaches would provide an important reference for interpreting task difficulty and assessing the significance of agent performance. Without such baselines, it remains unclear how much of the observed difficulty is specific to LLM-based hierarchical reasoning rather than the underlying combinatorial optimization problem. I believe if you dont use some traditional baseline like hybrid optimization to verify your concern/claim, the benchmark is not so meaningful. The proposed benchmark involves combinatorial optimization problems closely related to known NP-hard problems, like equipment allocation, spatial packing, and logistics network design. Please show the shortcoming/strength with quantifiable results of LLM/Agent of this problem, the paper will be better.

Questions for authors
  1. Have the authors independently verified the feasibility of all 32 tasks? In particular, can reference designs or feasibility evidence be provided for the 12 tasks that none of the evaluated agents solved?

  2. How would the evaluated agents compare with conventional heuristic or constraint-based optimization methods under the same task specifications?

  3. How does the benchmark distinguish hierarchical design reasoning from general spatial placement and constraint satisfaction? Is there evidence that the explicit subsystem structure is necessary for solving the tasks?

Overall, I find the proposed platform and problem formulation interesting and potentially valuable. However, without establishing the feasibility of all benchmark tasks or providing comparisons with conventional optimization methods, it is difficult to assess the broader significance of the benchmark. Otherwise, the study risks remaining a relatively narrow testbed for LLM performance rather than offering deeper insights into hierarchical design and optimization. I would encourage the authors to strengthen the experimental validation and include reproducible optimization baselines, so that the benchmark can provide meaningful insights to the operations research and engineering communities, rather than simply demonstrating how well LLMs perform on these tasks. I think it will take some time, but it is worthy.

1 comment
AuthorsOpenReview ↗

Response to Reviewer oY6y

[Q1] Have all 32 tasks been independently verified as feasible, particularly the 12 tasks unsolved by the evaluated models?

[A1] All 32 tasks now have verified successful reference designs. Please see our General Response section "[CQ3] Feasibility of All 32 Tasks" and the task-by-task summary in Appendix B.2 of the revised manuscript.

[Q2] How do the evaluated agents compare with conventional heuristic, constraint-based, or hybrid optimization methods under the same task specifications? Such baselines would help distinguish limitations of LLM agents from the difficulty of the underlying combinatorial optimization problem.

[A2] Thank you for suggesting a planning and optimization baseline. We added an end-to-end non-LLM method that met the production-rate criterion on 28 of the 32 tasks (87.5% task coverage), providing a concrete planning and optimization reference for interpreting agent performance. It combines mixed-integer linear programming (MILP) with domain-specific construction heuristics.

  • The MILP selects production-equipment types and integer counts to minimize their cost subject to material-balance and capacity constraints derived from recipes and the target production rate.

  • Reusable production-unit templates, greedy placement, and logistics and power construction rules then generate a complete factory.

The solver makes no LLM calls and does not read human reference designs or prior model-generated layouts.

After development and trial runs on the released suite, we froze the shared rules and parameters. The final evaluation generated and executed one design per task under the original task specifications and success criterion. All 32 designs passed static validation, executed successfully, and produced the target item. In the original experiments, DeepSeek V4 Flash, V4 Pro, GPT-5 mini, and Gemini 3 Flash Preview solved 19, 17, 12, and 12 tasks, respectively, counting a task as solved if any of its three independent runs succeeded. These task-coverage figures use the respective attempt counts stated above and do not constitute a controlled comparison. Code, designs, and evaluation records are available in the anonymous baseline package.

The method encodes machine-inserter-chest interfaces, construction order, and infrastructure rules, but still computes equipment quantities and placements for each task. It also uses explicit environment parameters, including the mining-productivity bonus, so its prior knowledge and information conditions differ from the original model experiments. MILP optimality applies only to the capacity subproblem, not the complete factory.

The comparison does not isolate hierarchical reasoning as the source of performance differences or separate the contributions of domain priors and individual algorithmic components.

[Q3] How does the benchmark distinguish hierarchical design reasoning from general spatial placement and constraint satisfaction? Is the explicit subsystem structure necessary for solving the tasks?

[A3] The production-rate requirement is one of our task constraints, so satisfying all task constraints constitutes success. Meeting them requires coordinating production dependencies, capacities, material interfaces, and shared infrastructure across the complete factory. A particular subsystem decomposition is not required or separately rewarded. Please see the General Response section "[CQ1] How Does the Benchmark Evaluate Hierarchical Design Synthesis, and What Is the Role of Explicit Subsystems?" for the distinction between placement validity and complete-task success, the interpretation of constraint satisfaction, and the role of explicit subsystem structure.

Reviewer_C4uiOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a Factorio-based benchmark for evaluating LLM agents on hierarchical design synthesis. It focuses on generating complete factory designs that coordinate production, spatial layout, logistics, and shared infrastructure under system-level constraints. Its main contribution is an evaluation framework combining declarative design submissions, static validation, and simulation-based performance measurement. Experiments with multiple LLMs demonstrate challenges in producing feasible factories that meet throughput requirements, while difficulty and cost analyses further characterize performance across tasks and models.

Critical strengths
  1. Meaningful coupling of design decisions and executable evaluation. The tasks require coordinating production dependencies, equipment capacity, spatial placement, logistics, and power. Measuring actual throughput provides a stronger functional criterion than checking design syntax or static feasibility alone. The separation between target production and target-rate success exposes useful differences in model behavior (Table 4).
  2. Careful reporting of difficulty and cost analyses. The difficulty profile is computed without model outcomes, and the appendix includes sensitivity checks, product-family bootstrap intervals, and controlled rate/map comparisons. The success-aware cost profile retains unsolved tasks in its denominator, while common-support comparisons complement it. These choices make the descriptive results more informative than a single success-rate ranking.
Critical weaknesses
  1. The validity and distinct contribution of the benchmark need stronger evidence. The paper does not provide verified reference solutions for all tasks, leaving uncertainty about the feasibility of the 12 tasks unsolved by every evaluated model. Moreover, although the tasks involve decisions at multiple levels, the evaluation does not establish how specifically it measures hierarchical organization rather than large-scale entity placement and complete-JSON generation. The comparison with the existing Factorio Learning Environment is primarily conceptual; a controlled comparison or representation/planning ablation would better establish the value added by this protocol (Sections 2.2–3.3).
  2. The diagnostic conclusions exceed the explanatory evidence. Raw violation counts are not directly comparable across rules: overlaps count entity pairs, some diagnostics are capped or summarized, and designs with no poles receive only one unpowered-consumer violation (Appendix B.2). Larger factories also offer more opportunities for violations. These counts identify recurring errors but do not establish the dominant causes of failure, particularly for statically valid factories with insufficient throughput. Likewise, the difficulty index strongly correlates with success, but capacity, logistics, and power are highly correlated, and power alone has a stronger marginal correlation than the aggregate index. Comparisons with simple scale-based predictors and runtime bottleneck analysis would strengthen the claimed diagnostic value.
Questions for authors
  1. How was feasibility established for the tasks, particularly the 12 unsolved by all evaluated models? Are verified successful designs or other feasibility checks available? This would help distinguish agent limitations from potentially infeasible task specifications.
  2. What evidence supports interpreting performance as a measure of hierarchical design synthesis, beyond entity placement and complete-design serialization? Please clarify which aspects of hierarchical organization are required or evaluated, and how this differs from existing Factorio evaluation protocols.
1 comment
AuthorsOpenReview ↗

Response to Reviewer C4ui

[Q1] How was feasibility established for all tasks, particularly the 12 tasks unsolved by every evaluated model?

[A1] All 32 tasks now have verified successful reference designs. Please see our General Response section "[CQ3] Feasibility of All 32 Tasks" and Appendix B.2 of the revised manuscript.

[Q2] What evidence shows that the benchmark evaluates hierarchical design synthesis beyond entity placement and complete-JSON generation? Which aspects of hierarchical organization are evaluated, and how does the protocol differ from FLE? A controlled comparison or representation/planning ablation would strengthen the evidence.

[A2] Our evaluation concerns completing hierarchical design tasks. Success requires coordinating component choices, production-stage capacities, material interfaces, and shared infrastructure to meet the factory's production target. Valid JSON and compliance with placement rules alone do not establish success. Please see the General Response section “[CQ1] How Does the Benchmark Evaluate Hierarchical Design Synthesis, and What Is the Role of Explicit Subsystems?” for a detailed explanation and examples.

We have expanded “[CQ2] Comparison with FLE” in our General Response to clarify two differences. Our evaluator instantiates complete submitted designs, so agents need not write construction programs or manage game operations such as character movement and item transfers between the character’s inventory and factory entities. This removes procedural implementation requirements that can confound the assessment of factory-design ability. We also rebuild each submission under standardized initial conditions, so its measured throughput cannot rely on inventories accumulated under earlier factory configurations. Together, these choices focus evaluation more directly on agents’ ability to produce complete designs that satisfy the task’s feasibility and production-rate requirements.

[Q3] Raw violation counts are not comparable across rules or factory sizes and do not establish the dominant causes of failure, especially throughput shortfalls. Capacity, logistics, and power are highly correlated, and power alone correlates more strongly with success than the aggregate difficulty index. Comparisons with simple scale-based predictors and runtime bottleneck analysis would strengthen the diagnostic claims.

[A3] We agree that raw violation counts differ in their units and reporting limits. We have updated the paper to compare reported static-rule violations, counting each rule at most once per submission. Entity overlap and unpowered consumers remain the two most frequently reported violations, occurring in 72.4% and 34.9% of all 8,409 evaluated submissions, respectively. These frequencies characterize recurring static-validation failures at the submission level, without adjustment for factory size. In future work, we will investigate why statically valid designs fail to reach the target production rate.

The five-dimensional profile is intended to characterize different design demands. We do not claim that it is the best predictor of model success. The observed associations may depend on the range of tasks covered, and some feature definitions or scoring choices may also need refinement. We will explore these possibilities in future work. The strong associations observed across models nevertheless support its current value as a specification-based proxy for relative task difficulty that can be computed before running any agent.

Reviewer_2u8fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Factorio Design Benchmark, a set of 32 factory design tasks covering eight products, with two production rates and two map settings for each product. Agents submit complete JSON designs. Designs that pass 22 static checks are built and run in the game to measure their production rates. Experiments with four models show that 73.7% of runs produce the target item, but only 40.9% reach the required rate within the allowed tolerance. The paper also introduces a five-dimensional difficulty measure based on task specifications and compares the equipment costs of successful designs.

Critical strengths
  1. The evaluation distinguishes between getting a factory to run and meeting its production requirements. Static checks and in-game tests provide objective evidence. Table 4 shows a clear gap between producing the target item and reaching the required rate, suggesting that checking functionality alone would miss important problems. The cost evaluation also considers both the cost of successful designs and the number of tasks solved.

  2. The task setup also makes it easier to compare different design requirements. Using different production rates and maps for the same product makes comparisons easier. The difficulty measure is built without using model results. Appendix B.2 also checks the findings across models, product families, and choices of weights. These results suggest that the measure is useful for comparing task difficulty.

Critical weaknesses
  1. The added value over prior work needs to be clearer. Factorio Learning Environment (FLE) [1] already includes tasks that require complete automated production lines and target production rates. This paper changes the design submission format, validation rules, and task settings, but provides little direct evidence of which additional hierarchical design abilities these changes test. In particular, blocks appear to serve mainly as groups of entities organized by location, and their role in execution and scoring is not clear.

  2. The hardest tasks lack verified reference designs. None of the models solved 12 of the 32 tasks, but the paper does not report working designs for these tasks under the same rules. Without reference designs, it is unclear whether these failures are due to the agents or the difficulty of the tasks themselves.

  3. The initial conditions are not fully described. The paper does not give enough detail about initial fuel, material inventories, or research bonuses. The complete example in the appendix also contains details that need explanation. These settings matter for understanding the reported production rates and should be stated clearly.

[1] Hopkins J, Bakler M, Khan A. Factorio learning environment[J]. Advances in Neural Information Processing Systems, 2026, 38.

Questions for authors
  1. Do all 32 tasks have known working designs that pass the same static checks and in-game evaluation? Could you provide these designs, especially for the 12 tasks that none of the models solved?

  2. Can a valid factory place all its production equipment in a single block? If so, what does the explicit block structure add to the benchmark’s assessment of design skills?

  3. What initial fuel, material inventories, and research bonuses are provided when a factory is created? The complete example in Appendix A.3.1 uses one mining drill and reports 37.42 iron plates per minute, but lists no equipment for supplying fuel to the boiler. Could you explain how these details relate to the settings used for this run?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2u8f

[Q1] FLE already includes automated production lines and target production rates. Which additional hierarchical design abilities does this benchmark assess? What role do blocks play in execution and scoring, and what do they add if all production equipment can be placed in a single block?

[A1] We have expanded “[CQ2] Comparison with FLE” in our General Response to clarify two differences. Our evaluator instantiates complete submitted designs, so agents need not write construction programs or manage game operations such as character movement and item transfers between the character’s inventory and factory entities. This removes procedural implementation requirements that can confound the assessment of factory-design ability. We also rebuild each submission under standardized initial conditions, so its measured throughput cannot rely on inventories accumulated under earlier factory configurations. Together, these choices focus evaluation more directly on agents’ ability to produce complete designs that satisfy the task’s feasibility and production-rate requirements.

Blocks organize equipment, local layouts, and material interfaces. Their boundaries, entity containment, and local interfaces are checked during static validation, but blocks are not game entities and receive no separate score. Multiple blocks are not required. A single-block design must still coordinate production stages and shared infrastructure to meet the task requirements. Please see the General Response sections “[CQ1] How Does the Benchmark Evaluate Hierarchical Design Synthesis, and What Is the Role of Explicit Subsystems?” and “[CQ2] Comparison with FLE” for a detailed explanation and examples.

[Q2] Do all 32 tasks have working designs that pass the same static checks and in-game evaluation, especially the 12 tasks unsolved by the evaluated models?

[A2] Yes. All 32 tasks now have verified successful reference designs. Please see our General Response section "[CQ3] Feasibility of All 32 Tasks" and the task-by-task summary in Appendix B.2 of the revised manuscript.

[Q3] What initial conditions, including fuel, material inventories, and research bonuses, are provided across all tasks? How does the appendix design produce 37.42 iron plates/min with one mining drill and no boiler-fueling equipment?

[A3] Thank you for requesting these details.

  • Each design is instantiated in a reset environment without preloaded ores or intermediate products.

  • Boiler coal is supplied automatically during setup and replenished throughout warm-up and monitoring, so no boiler-fueling equipment is required.

  • All tasks start with technologies researched and the corresponding production and logistics bonuses enabled.

The 40% mining-productivity bonus raises a single electric mining drill's theoretical iron-ore output from 30 to 42 items/min. Since iron smelting consumes one ore per plate, this supply can support the example's measured 37.42 plates/min. The shared initialization settings, research bonuses, and measurement procedure are now documented in Appendix B.1 of the revised manuscript. Additional evaluation settings and provenance are available in the anonymous evaluation-configuration record.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response(1/2)

[CQ1] How Does the Benchmark Evaluate Hierarchical Design Synthesis, and What Is the Role of Explicit Subsystems? [2u8f Q1, C4ui Q2, oY6y Q3]

Our aim is to evaluate agents' ability to complete hierarchical design tasks. These tasks require selecting component types and quantities, specifying their positions and configurations, and organizing them into interacting subsystems that form a complete system satisfying constraints and performance objectives. In our factory-design tasks, this entails equipment selection and configuration, capacity allocation across production stages, material-interface design, and coordination of factory-wide logistics and power. Equipment and local transfer interfaces form mining, smelting, and assembly units, which together constitute a factory through material dependencies and shared infrastructure. Conversely, the factory's production target determines the capacities needed at each stage, affecting equipment quantities and interface configurations. For example, adding final-stage assemblers to raise throughput also increases upstream material, transport, and power requirements. These choices must jointly support the production target, with overall feasibility and performance assessed on the complete design.

We evaluate whether decisions at these levels jointly produce a system that meets the task requirements, using static checks and measured throughput. Static checks assess necessary conditions for production-chain completeness, material interfaces, spatial layout, and infrastructure consistency. Designs that pass these checks are instantiated in Factorio, and their production rates are monitored after warm-up to assess target-rate success. Regarding the reviewers' question about spatial placement and constraint satisfaction, valid JSON and a layout obeying Factorio's placement rules do not guarantee success. For example, a requester chest supplying an intermediate production stage may request excessive amounts of a shared raw material, leaving other stages short and preventing the factory from reaching its target throughput despite a valid layout. Addressing this requires hierarchical design capabilities including production-dependency analysis, capacity allocation, line balancing, and coordination of shared resources. The production-rate requirement is itself a system-level constraint. Meeting it requires coordinating capacities and material flows across production stages, beyond satisfying the static placement and interface checks.

Our design representation supports this component-to-system organization through production blocks and global entities for shared infrastructure. Blocks explicitly group equipment, local layouts, and material interfaces. Static validation checks their boundaries, entity containment, and local interface consistency. After validation, the equipment is instantiated in the game. Blocks themselves are not game entities and do not participate in simulation. We assign no separate reward to block count or a particular decomposition, and multiple blocks are not required. All production equipment may be placed in one block, provided the design passes the checks. Functional dependencies between production stages remain, so success still requires coordinating these stages and their shared infrastructure.

[CQ3] Feasibility of All 32 Tasks [2u8f Q2, C4ui Q1, oY6y Q1]

We thank the reviewers for raising the question of task feasibility. We have now completed and verified reference designs for all 12 tasks that none of the evaluated models solved. Together with successful model-generated designs for the other 20 tasks, these provide one successful reference design for every task in the 32-task suite. All designs pass static validation, execute successfully, and meet the production-rate acceptance criterion under the original task specifications and evaluation settings.

The complete design JSON files and evaluation records are provided in the anonymous reference-design package. We have added a task-by-task summary to Appendix B.2, "Reference Designs and Task Feasibility," of the revised manuscript. These reference designs do not alter the reported model results.

AuthorsOpenReview ↗

General Response (2/2)

[CQ2] Comparison with FLE [2u8f Q1, C4ui Q2]

Our benchmark and the Factorio Learning Environment (FLE) lab-play planning tasks both require automated production lines that meet target rates. The key differences are that (1) agents submit complete production-line specifications for the evaluator to instantiate, without having to write construction programs or control the game character, and (2) each submitted design is evaluated independently of material inventories accumulated during earlier interactions. These choices focus the assessment on the hierarchical design of complete production lines and their operation under common starting conditions.

1. Separating production-line design from procedural construction.

FLE’s tools automate some placement and connection decisions while retaining operations for game character control and inventory management. To construct their intended factories, agents must coordinate these tool calls through Python programs that modify the game state. For example, FLE exposes move_to, and its place_entity implementation rejects targets beyond the character’s reach and requires moving closer. Even with equipment and positions already determined, the agent must arrange this movement before placement. These procedural requirements introduce additional failure modes that can confound the assessment of factory-design ability. An agent may have an otherwise feasible factory design yet fail to construct it because of errors in its program or game operations. FLE’s human baseline illustrates this additional programming burden: it recorded 87 errors, including 57 programmatic errors such as incorrect variable references or API usage (Appendix H.7).

In our benchmark, each submission explicitly specifies a complete production-line design in JSON, including equipment configurations, positions, and orientations, which together determine material-transfer interfaces and infrastructure connections. The evaluator handles its instantiation and configuration. Agents can iteratively revise their designs using evaluation feedback, but each revision is submitted as a complete candidate specification.

Delegating construction to the evaluator allows agents to focus on equipment choices, capacity allocation, layout, and subsystem coordination. The resulting evaluation more directly assesses whether their design decisions produce a feasible factory meeting the production-rate criterion.

2. Evaluating each design independently of construction history.

FLE evaluates new target production over finite measurement windows while retaining the evolving game state across interactions. Measured output can therefore depend on both the current layout and inventories accumulated during earlier steps. For example, a layout could temporarily meet the target by consuming stored intermediates even after their production equipment has been removed. Such a measurement would not by itself establish that the current layout could achieve the same throughput when rebuilt from standardized initial conditions.

We rebuild each submitted design under standardized initial conditions before warm-up and monitoring. Material inventories left by earlier candidates therefore cannot contribute to its measured output. A successful submission establishes that the specified design meets the production-rate criterion under the same initialization and monitoring protocol used for other candidates.

The complete specification makes each candidate reconstructible, and standardized reconstruction makes its throughput evaluation independent of inventories accumulated during earlier revisions. These two choices work together: the agent submits a complete factory configuration, and the evaluator checks its operation from a common starting point.

Beyond these core differences, our suite crosses two target rates with two maps per product to assess adaptation to capacity and site requirements. We also compare total entity cost among successful designs for the same task, distinguishing factories that meet the same production target using different amounts of equipment. These provide supplementary dimensions for assessing design adaptability and efficiency.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.