pith. sign in

arxiv: 2607.02082 · v1 · pith:NFYWRWLQnew · submitted 2026-07-02 · 💻 cs.NE · cs.AI

Evolutionary Wave Function Collapse

Dipika Rajesh , Ahmed Khalifa , Julian Togelius This is my paper

Pith reviewed 2026-07-03 03:11 UTC · model grok-4.3

classification 💻 cs.NE cs.AI
keywords wave function collapseevolutionary optimizationprocedural content generationmaze generationdungeon generationgenotype-phenotype mappinglocal constraints
0
0 comments X

The pith

Evolving the small input examples for Wave Function Collapse improves generated level quality when properties emerge from local relationships.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper explores combining evolutionary search with Wave Function Collapse by optimizing the small example inputs that WFC learns from, rather than evolving complete levels directly. WFC functions as the mapping from those evolved inputs to larger generated outputs that get scored by domain fitness functions. Experiments compare performance on maze connectivity maps, where local rules drive connectivity, against Zelda-style dungeons that need global layout properties. The approach yields better results precisely when objectives match what local adjacency constraints can produce. Readers would care because WFC is already common in procedural content generation, and this method offers a way to steer it without redesigning the core algorithm.

Core claim

Evolutionary optimization over the small input examples used by WFC improves the quality of generated levels in domains where properties emerge from local relationships, while domains requiring global constraints remain challenging. WFC acts as the genotype-to-phenotype mapping, and generated levels are evaluated through domain-specific fitness functions.

What carries the argument

Evolutionary optimization of WFC input examples, with WFC serving as the genotype-to-phenotype mapping.

If this is right

  • Evolutionary search guides WFC effectively when target objectives align with local structure.
  • Generation quality improves for connectivity maps such as mazes.
  • Global constraint domains such as Zelda layouts stay difficult.
  • The method supplies an alternative to direct evolution of complete levels in suitable cases.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same input-evolution strategy could be tested on other local-structure PCG tasks such as tile-based terrain.
  • Hybrid approaches that add explicit global checks after WFC expansion might extend the method to harder domains.
  • Because inputs stay small, search remains tractable even when target outputs grow large.

Load-bearing premise

The fitness functions accurately capture intended quality metrics and WFC's local constraints suffice to express target structures once inputs are optimized.

What would settle it

An experiment in the maze domain finding no statistically significant quality gain from evolved inputs over fixed or random inputs would falsify the reported improvement.

Figures

Figures reproduced from arXiv: 2607.02082 by Ahmed Khalifa, Dipika Rajesh, Julian Togelius.

Figure 1
Figure 1. Figure 1: Outputs generated using evolutionary and random search in the Maze and Zelda domains respectively. Evolutionary [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Best fitness over generations in the Maze domain. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
read the original abstract

Wave Function Collapse (WFC) is a widely used procedural content generation method that learns local adjacency constraints from example inputs to generate larger outputs. In this paper, we explore combining WFC with evolutionary search by evolving the small input examples used by WFC rather than directly evolving complete levels. In this approach, WFC acts as a genotype-to-phenotype mapping. The generated levels are then evaluated through domain-specific fitness functions. We evaluate the method in two domains with different relationships between local and global structure: Maze connectivity maps and Zelda-style dungeon layouts. Our results show that evolutionary optimization over WFC inputs improves generation quality in domains where properties emerge from local relationships, while domains requiring global constraints remain challenging. These findings suggest that evolutionary search can effectively guide WFC generation when target objectives align with local structure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes evolving the small input examples fed to Wave Function Collapse (WFC) rather than evolving complete levels directly, treating WFC as a genotype-to-phenotype mapping whose outputs are scored by domain-specific fitness functions. Experiments are reported in two domains (maze connectivity maps and Zelda-style dungeon layouts) with the central claim that evolutionary optimization over WFC inputs improves generation quality when target properties emerge from local relationships, while domains requiring global constraints remain challenging.

Significance. If the empirical results are robust, the work would supply a concrete hybrid PCG technique that exploits WFC's local constraint learning while using evolution to steer the choice of training examples, together with an empirical distinction between locally emergent and globally constrained objectives.

major comments (2)
  1. [Abstract] Abstract: the claim that evolutionary optimization 'improves generation quality' is presented without any description of experimental protocol (population size, number of generations, number of independent runs, baselines, or statistical tests), rendering the local-vs-global distinction impossible to evaluate from the provided information.
  2. [Method / Evaluation] Fitness functions (implicit in the method description): the central claim requires that the domain-specific fitness functions correctly measure the intended properties (connectivity for mazes, layout quality for Zelda). No definitions, validation, or discussion of possible proxy failures or global violations are supplied, so observed differences could be artifacts of the fitness rather than evidence about WFC's local-constraint limitation.
minor comments (1)
  1. [Abstract] The abstract and method overview should explicitly state the evolutionary algorithm parameters and the precise form of the fitness functions used in each domain.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on the abstract and fitness functions. We address each major comment below and will revise the manuscript accordingly to improve clarity and support for the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that evolutionary optimization 'improves generation quality' is presented without any description of experimental protocol (population size, number of generations, number of independent runs, baselines, or statistical tests), rendering the local-vs-global distinction impossible to evaluate from the provided information.

    Authors: The full experimental protocol, including population size, generations, independent runs, baselines (such as random WFC inputs), and statistical tests, is detailed in the Methods and Results sections. We agree the abstract would benefit from a brief summary of these elements to allow readers to evaluate the local-vs-global distinction more readily. We will revise the abstract to include a concise description of the protocol. revision: yes

  2. Referee: [Method / Evaluation] Fitness functions (implicit in the method description): the central claim requires that the domain-specific fitness functions correctly measure the intended properties (connectivity for mazes, layout quality for Zelda). No definitions, validation, or discussion of possible proxy failures or global violations are supplied, so observed differences could be artifacts of the fitness rather than evidence about WFC's local-constraint limitation.

    Authors: We agree that explicit definitions, validation, and discussion of limitations are needed to substantiate the claims. The manuscript describes the fitness functions in the evaluation section, but we will expand this with precise definitions (e.g., maze connectivity via connected-component analysis), validation against manual checks, and a discussion of potential proxy failures such as global violations not captured by local metrics. This will clarify that differences arise from alignment with local constraints. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical method comparison without derivations or self-referential predictions

full rationale

The paper presents an experimental combination of evolutionary search over WFC input examples (genotype-to-phenotype via WFC) evaluated by domain-specific fitness functions in maze and Zelda domains. No equations, parameter fittings, predictions, or uniqueness theorems are claimed. Results are direct empirical comparisons of generation quality, with no load-bearing self-citations or reductions of outputs to inputs by construction. This is a standard empirical PCG study and self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review yields minimal ledger entries; the method rests on standard assumptions of WFC and evolutionary algorithms without new fitted parameters or invented entities.

axioms (1)
  • domain assumption WFC learns local adjacency constraints from example inputs that can be evolved to improve output quality
    Central premise stated in the abstract description of the genotype-to-phenotype mapping.

pith-pipeline@v0.9.1-grok · 5655 in / 1036 out tokens · 28897 ms · 2026-07-03T03:11:51.232281+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references

  1. [1]

    Wavefunctioncollapse is constraint solving in the wild,

    I. Karth and A. M. Smith, “Wavefunctioncollapse is constraint solving in the wild,” inFoundations of Digital Games. ACM, 2017, pp. 1–10

    2017

  2. [2]

    Shaker, J

    N. Shaker, J. Togelius, and M. J. Nelson,Procedural content generation in games. Springer, 2016

    2016

  3. [3]

    Procedural content generation via machine learning (pcgml),

    A. Summerville, S. Snodgrass, M. Guzdial, C. Holmg ˚ard, A. K. Hoover, A. Isaksen, A. Nealen, and J. Togelius, “Procedural content generation via machine learning (pcgml),”Transactions on Games, vol. 10, no. 3, pp. 257–270, 2018

    2018

  4. [4]

    Wave function collapse in bad north,

    O. Stalberg, “Wave function collapse in bad north,” 2018, accessed: https://www.youtube.com/watch?v=0bcZb-SsnrA

    2018

  5. [5]

    Content reinjection for super metroid,

    R. Mawhorter, B. Aytemiz, I. Karth, and A. Smith, “Content reinjection for super metroid,” inArtificial Intelligence and Interactive Digital Entertainment, vol. 17, no. 1. AAAI, 2021, pp. 172–178

    2021

  6. [6]

    Enhancing wave function collapse with design-level constraints,

    A. Sandhu, Z. Chen, and J. McCoy, “Enhancing wave function collapse with design-level constraints,” inFoundations of Digital Games. ACM, 2019, pp. 1–9

    2019

  7. [7]

    Automatic generation of game levels based on controllable wave function collapse algorithm,

    D. Cheng, H. Han, and G. Fei, “Automatic generation of game levels based on controllable wave function collapse algorithm,” inInternational Conference on Entertainment Computing. Springer, 2020, pp. 37–50

    2020

  8. [8]

    Combining constructive procedural dungeon generation methods with wavefunctioncollapse in top-down 2d games,

    P. Minini and J. Assuncao, “Combining constructive procedural dungeon generation methods with wavefunctioncollapse in top-down 2d games,” inSBGames, 2020

    2020

  9. [9]

    Neurosymbolic map generation with vq-vae and wfc,

    I. Karth, B. Aytemiz, R. Mawhorter, and A. M. Smith, “Neurosymbolic map generation with vq-vae and wfc,” inFoundations of Digital Games. ACM, 2021, pp. 1–6

    2021

  10. [10]

    Hierarchical semantic wave function collapse,

    S. Alaka and R. Bidarra, “Hierarchical semantic wave function collapse,” inFoundations of Digital Games. ACM, 2023, pp. 1–10

    2023

  11. [11]

    Automatic generation of game content using a graph-based wave function collapse algorithm,

    H. Kim, S. Lee, H. Lee, T. Hahn, and S. Kang, “Automatic generation of game content using a graph-based wave function collapse algorithm,” inConference on Games. IEEE, 2019, pp. 1–4

    2019

  12. [12]

    Better resemblance without bigger patterns: Making context-sensitive decisions in wfc,

    B. Bateni, I. Karth, and A. Smith, “Better resemblance without bigger patterns: Making context-sensitive decisions in wfc,” inFoundations of Digital Games. ACM, 2023, pp. 1–11

    2023

  13. [13]

    Leveraging reinforcement learning and wavefunctioncollapse for improved procedural level generation,

    M. Babin and M. Katchabaw, “Leveraging reinforcement learning and wavefunctioncollapse for improved procedural level generation,” in Foundations of Digital Games. ACM, 2021, pp. 1–8

    2021

  14. [14]

    Genetic-wfc: Extending wave function collapse with genetic search,

    R. Bailly and G. Levieux, “Genetic-wfc: Extending wave function collapse with genetic search,”Transactions on Games, vol. 15, no. 1, pp. 36–45, 2022

    2022

  15. [15]

    Search- based procedural content generation: A taxonomy and survey,

    J. Togelius, G. N. Yannakakis, K. O. Stanley, and C. Browne, “Search- based procedural content generation: A taxonomy and survey,”Trans- actions on Computational Intelligence and AI in Games, vol. 3, no. 3, pp. 172–186, 2011

    2011

  16. [16]

    Ropossum: An authoring tool for designing, optimizing and solving cut the rope levels,

    N. Shaker, M. Shaker, and J. Togelius, “Ropossum: An authoring tool for designing, optimizing and solving cut the rope levels,” inArtificial Intelligence and Interactive Digital Entertainment, vol. 9, no. 1. AAAI, 2013, pp. 215–216

    2013

  17. [17]

    An experiment in automatic game design,

    J. Togelius and J. Schmidhuber, “An experiment in automatic game design,” inSymposium On Computational Intelligence and Games. IEEE, 2008, pp. 111–118

    2008

  18. [18]

    Metacompose: A compo- sitional evolutionary music composer,

    M. Scirea, J. Togelius, P. Eklund, and S. Risi, “Metacompose: A compo- sitional evolutionary music composer,” inEvolutionary and Biologically Inspired Music, Sound, Art and Design. Springer, 2016, pp. 202–217

    2016

  19. [19]

    A procedural procedural level generator generator,

    M. Kerssemakers, J. Tuxen, J. Togelius, and G. N. Yannakakis, “A procedural procedural level generator generator,” inConference on Computational Intelligence and Games. IEEE, 2012, pp. 335–341

    2012

  20. [20]

    Multi-objective level generator generation with marahel,

    A. Khalifa and J. Togelius, “Multi-objective level generator generation with marahel,” inFoundations of Digital Games. ACM, 2020, pp. 1–8

    2020

  21. [21]

    The procedural content generation benchmark: an open- source testbed for generative challenges in games,

    A. Khalifa, R. Gallotta, M. Barthet, A. Liapis, J. Togelius, and G. N. Yannakakis, “The procedural content generation benchmark: an open- source testbed for generative challenges in games,” inProceedings of the 20th International Conference on the Foundations of Digital Games, 2025, pp. 1–12

    2025