arxiv: 2605.01245 · v1 · submitted 2026-05-02 · 💻 cs.HC · cs.AI

Recognition: unknown

The Garden of Forking Paths: Narrative Arc-Conditioned Gameplay Planning

Chenliang Huang, Chun Ming Louis Po, Hangyu Zhou, Julian Togelius, Sam Earle, Timothy Merino, Yunge Wen, Zhuo Zeng

Authors on Pith no claims yet

Pith reviewed 2026-05-09 17:46 UTC · model grok-4.3

classification 💻 cs.HC cs.AI

keywords narrative arcbranching gameplayprocedural generationdungeon graphLLM game designmultimodal alignmentinteractive systemstory-conditioned planning

0 comments

The pith

Forking Garden generates branching games from user storylines by first creating independent gameplay nodes then assembling them into a dungeon graph using narrative arc constraints.

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

The paper targets the gap where procedural game generation ignores established narrative patterns such as the Hero's Journey or three-act structure. It introduces a two-stage process that first produces a diverse set of standalone nodes, each with aligned gameplay, visuals, and other elements, and then links those nodes into branching paths according to a chosen story arc. This separation allows the system to produce varied yet coherent dungeons from a simple user-provided storyline. The result is an interactive tool that outputs playable branching structures without requiring designers to hand-craft every branch. A sympathetic reader would care because it points toward more story-resonant procedural content that still feels authored rather than random.

Core claim

We propose Forking Garden, a framework for narrative arc-conditioned gameplay planning that generates branching games from user-provided storylines. Our approach first generates a diverse pool of independent nodes, then assembles them into a dungeon graph via arc-guided constraint algorithms, where each node achieves multimodal alignment of gameplay elements. We develop an end-to-end interactive system that instantiates the framework.

What carries the argument

Arc-guided constraint algorithms that take a pool of independently generated nodes and link them into a coherent branching dungeon graph while preserving the supplied narrative arc.

If this is right

Branching games can be produced directly from storylines while keeping explicit narrative structure across paths.
Each node in the graph carries internally consistent gameplay, visual, and other elements.
Diversity among branches is preserved because nodes are generated independently before assembly.
An interactive interface lets users supply the storyline and receive a ready-to-play dungeon graph.
The method reduces the amount of post-generation human adjustment needed for narrative coherence.

Where Pith is reading between the lines

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

The separation of node creation from graph assembly could be tested in genres other than dungeons, such as open-world or narrative-driven adventure games.
If the constraint stage works well, similar pipelines might improve coherence in other LLM-driven creative tools like interactive fiction or level design assistants.
Player studies could measure whether games produced this way feel more narratively satisfying than purely random procedural outputs.
The framework implicitly suggests that enforcing archetypal story shapes may be a practical way to steer generative systems toward culturally resonant results.

Load-bearing premise

Arc-guided constraint algorithms can reliably turn independently generated nodes into coherent branching structures that stay aligned with the narrative arc and keep multimodal consistency without extra human fixes.

What would settle it

Feed the system several user storylines, generate the graphs, and check whether the resulting paths consistently follow the intended arc, maintain node-to-node coherence, and require no manual repair to become playable.

Figures

Figures reproduced from arXiv: 2605.01245 by Chenliang Huang, Chun Ming Louis Po, Hangyu Zhou, Julian Togelius, Sam Earle, Timothy Merino, Yunge Wen, Zhuo Zeng.

**Figure 1.** Figure 1: Overview of the gameplay planning pipeline. Given a storyline and protagonist (e.g., Little Red Riding Hood), the view at source ↗

**Figure 2.** Figure 2: End-to-end system for narrative arc-conditioned game generation. (A) Users input a storyline and protagonist. (B) The view at source ↗

**Figure 3.** Figure 3: Dungeon level example for Unity integration. view at source ↗

**Figure 4.** Figure 4: Example of a Rise-Fall narrative archetype view at source ↗

read the original abstract

Narrative archetypes (e.g., Hero's Journey, Three-act structure) provide universal story structures that resonate across cultures and media and are important for video game storytelling, yet existing LLM-based methods lack explicit use of these archetypes in procedurally generated games. We propose Forking Garden, a framework for narrative arc-conditioned gameplay planning that generates branching games from user-provided storylines. Our approach first generates a diverse pool of independent nodes, then assembles them into a dungeon graph via arc-guided constraint algorithms, where each node achieves multimodal alignment of gameplay elements. We develop an end-to-end interactive system that instantiates the framework.

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.

Desk Editor's Note private letter to a colleague

The paper sketches a two-stage pipeline for arc-conditioned branching game graphs but offers no tests of whether the constraint assembly actually works at scale.

read the letter

The main contribution is the split between generating a diverse pool of independent nodes first and then assembling them into a dungeon graph guided by narrative arc constraints. This is a clean way to keep generation flexible while trying to enforce story structure like the Hero's Journey or three-act arcs, and the multimodal alignment per node is a sensible target for gameplay elements. They also describe an end-to-end interactive system that takes user storylines as input, which makes the idea feel closer to a usable tool than pure theory. That pipeline is new enough in the combination to stand out from generic LLM dungeon generators. The soft spot is exactly where the stress test points: the arc-guided constraint algorithms are presented as the mechanism that delivers coherent, aligned graphs without extra fixes, yet the paper gives no success rates, no examples of solved versus failed assemblies, and no ablation on what happens when nodes are incompatible. Without that evidence it is hard to know whether the approach stays end-to-end or quietly relies on human cleanup for realistic inputs. This is aimed at people working on procedural storytelling and AI-assisted game design who already think about narrative archetypes. A reader could borrow the two-stage structure as a starting template even if they end up replacing the constraint solver. I would send it to peer review if the authors add at least preliminary validation or failure cases on the assembly step, because the framing is clear and the gap it targets is real, but the current version is still a proposal rather than a demonstrated method.

Referee Report

2 major / 1 minor

Summary. The paper proposes Forking Garden, a framework for narrative arc-conditioned gameplay planning that generates branching dungeon games from user-provided storylines. It first creates a diverse pool of independent nodes, then assembles them into graphs via arc-guided constraint algorithms that enforce multimodal alignment of gameplay elements with narrative archetypes (e.g., Hero's Journey). An end-to-end interactive system instantiates the framework.

Significance. If the assembly step can be shown to work reliably, the separation of flexible LLM node generation from hard arc-guided constraints offers a principled way to incorporate universal narrative structures into procedural content generation, a gap in current LLM-based game design tools. The conceptual distinction between node creation and arc-conditioned assembly is a clear strength and could support reproducible, archetype-driven branching narratives.

major comments (2)

Abstract and Proposed Framework: The central claim that arc-guided constraint algorithms can reliably assemble independently generated nodes into coherent branching graphs while preserving narrative alignment and multimodal consistency is load-bearing, yet the manuscript supplies no formal invariants, success-rate bounds, compatibility checks, or even a single worked example of a valid assembly for a realistic storyline. Without these, the 'end-to-end without additional intervention' property remains unverified.
Abstract: No experiments, quantitative metrics, ablation studies, error analysis, or user validation of the generated graphs are presented. This absence directly undermines assessment of whether the constraint solver routinely finds solutions or whether multimodal alignment holds in practice.

minor comments (1)

Abstract: The term 'multimodal alignment' is used without specifying the modalities (visual, mechanical, audio, etc.) or the alignment metric, which should be defined early for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important areas for strengthening the presentation and validation of the Forking Garden framework. We agree that additional illustrative examples and empirical details will improve the manuscript and plan targeted revisions accordingly.

read point-by-point responses

Referee: Abstract and Proposed Framework: The central claim that arc-guided constraint algorithms can reliably assemble independently generated nodes into coherent branching graphs while preserving narrative alignment and multimodal consistency is load-bearing, yet the manuscript supplies no formal invariants, success-rate bounds, compatibility checks, or even a single worked example of a valid assembly for a realistic storyline. Without these, the 'end-to-end without additional intervention' property remains unverified.

Authors: We acknowledge that the manuscript does not include formal invariants, success-rate bounds, or a worked example of the assembly process. The framework separates node generation (via LLM) from assembly (via arc-guided constraints), where the constraints are designed to enforce narrative archetype alignment and multimodal consistency by construction during graph construction. In the revised version, we will add a detailed worked example showing node pool creation and constraint-based assembly for a realistic storyline (e.g., a Hero's Journey arc), including the specific constraints applied and the resulting graph. We will also describe the constraint types and their guarantees more explicitly, though deriving full formal invariants or probabilistic bounds would require further theoretical analysis that we can outline as future work. revision: partial
Referee: Abstract: No experiments, quantitative metrics, ablation studies, error analysis, or user validation of the generated graphs are presented. This absence directly undermines assessment of whether the constraint solver routinely finds solutions or whether multimodal alignment holds in practice.

Authors: The initial submission focuses on the conceptual framework and end-to-end system implementation rather than extensive evaluation. We agree that quantitative assessment is necessary to demonstrate reliability. In the revision, we will add a new evaluation section reporting: success rates of the constraint solver on a set of sample storylines, qualitative analysis of multimodal alignment in generated graphs, and an error analysis of failure cases (e.g., when no valid assembly exists). Ablation studies on the constraint components can also be included. Full user validation studies are valuable but may be noted as planned future work given the scope of this paper. revision: yes

Circularity Check

0 steps flagged

No circularity: framework proposal with no derivations or self-referential reductions

full rationale

The paper presents Forking Garden as a high-level system architecture: independent node generation followed by assembly via arc-guided constraints into dungeon graphs. No equations, fitted parameters, predictions, or first-principles derivations appear in the provided text. The central claims are engineering proposals rather than mathematical reductions, and no self-citations are invoked as load-bearing uniqueness theorems or ansatzes. The derivation chain is therefore self-contained with no steps that reduce to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on the background assumption that narrative archetypes are effective guides for gameplay alignment and on the unproven effectiveness of the newly proposed constraint algorithms.

axioms (1)

domain assumption Narrative archetypes (e.g., Hero's Journey, Three-act structure) provide universal story structures that resonate across cultures and media and are important for video game storytelling.
Stated directly in the abstract as the motivation for the work.

invented entities (2)

Forking Garden framework no independent evidence
purpose: Narrative arc-conditioned gameplay planning that generates branching games from user-provided storylines
Newly proposed end-to-end system.
arc-guided constraint algorithms no independent evidence
purpose: Assemble independent nodes into a dungeon graph while achieving multimodal alignment
Core technical component of the proposed approach.

pith-pipeline@v0.9.0 · 5418 in / 1325 out tokens · 50494 ms · 2026-05-09T17:46:53.538863+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 12 canonical work pages

[1]

2008.The Hero with a Thousand Faces(3rd ed.)

Joseph Campbell. 2008.The Hero with a Thousand Faces(3rd ed.). Joseph Campbell Foundation; New World Library. Originally published in 1949. CHI EA ’26, April 13–17, 2026, Barcelona, Spain Wen, Huang, Zhou, Zeng, Po, et al

2008
[2]

Jing Chen, Xinyu Zhu, Cheng Yang, Chufan Shi, Yadong Xi, Yuxiang Zhang, Junjie Wang, Jiashu Pu, Rongsheng Zhang, Yujiu Yang, et al. 2024. Hollmwood: Unleashing the Creativity of Large Language Models in Screenwriting via Role Playing.arXiv preprintarXiv:2406.11683 (2024)

work page arXiv 2024
[3]

Sam Earle, Samyak Parajuli, and Andrzej Banburski-Fahey. 2025. DreamGarden: A Designer Assistant for Growing Games from a Single Prompt. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 57, 19 pages. doi:10.1145/3706598.3714233

work page doi:10.1145/3706598.3714233 2025
[4]

Carlos Gómez-Rodríguez and Paul Williams. 2023. A Confederacy of Mod- els: a Comprehensive Evaluation of LLMs on Creative Writing.arXiv preprint arXiv:2310.08433(2023)

work page arXiv 2023
[5]

Senyu Han, Lu Chen, Li-Min Lin, Zhengshan Xu, and Kai Yu. 2024. Ibsen: Director- Actor Agent Collaboration for Controllable and Interactive Drama Script Gener- ation.arXiv preprintarXiv:2407.01093 (2024)

work page arXiv 2024
[6]

Sergio Poo Hernandez, Vadim Bulitko, and Marcia Spetch. 2015. Keeping the Player on an Emotional Trajectory in Interactive Storytelling. InProceedings of the Eleventh Artificial Intelligence and Interactive Digital Entertainment Conference (AIIDE ’15). Association for the Advancement of Artificial Intelligence. doi:10. 1609/aiide.v11i1.12783

2015
[7]

Shuo Huang. 2025. Word2Minecraft: Generating 3D Game Levels through Large Language Models.arXiv preprint(2025). https://arxiv.org/abs/2503.16536 New York University Tandon School of Engineering

work page arXiv 2025
[8]

Mete Ismayilzada, Claire Stevenson, and Lonneke van der Plas. 2024. Evaluating Creative Short Story Generation in Humans and Large Language Models.arXiv preprint arXiv:2411.02316(2024)

work page arXiv 2024
[9]

Vignesh Kumaran, Jonathan Rowe, and James Lester. 2024. NarrativeGenie: Gen- erating Narrative Beats and Dynamic Storytelling with Large Language Models. InProceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE), Vol. 20. 76–86. doi:10.1609/aiide.v20i1.31868

work page doi:10.1609/aiide.v20i1.31868 2024
[10]

Jorge Leandro, Sudha Rao, Michael Xu, Weijia Xu, Nebojsa Jojic, Chris Brockett, and Bill Dolan. 2024. GENEVA: GENErating and Visualizing Branching Narratives Using LLMs. InIEEE Conference on Games 2024

2024
[11]

2011.Interactive Storytelling for Video Games (1st ed.)

Josiah Lebowitz and Chris Klug. 2011.Interactive Storytelling for Video Games (1st ed.). Focal Press, Burlington, MA, USA

2011
[12]

Bernstein

Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simu- lacra of Human Behavior. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST ’23). Association for Computing Machinery, New York, NY, USA, 1–22

2023
[13]

1928.Morphology of the Folktale(second ed.)

Vladimir Propp. 1928.Morphology of the Folktale(second ed.). Publications of the American Folklore Society, Vol. 9. University of Texas Press, Austin, TX. Originally published in Russian in 1928

1928
[14]

A. J. Reagan, L. Mitchell, D. Kiley, C. M. Danforth, and P. S. Dodds. 2016. The emotional arcs of stories are dominated by six basic shapes.EPJ Data Science5, 31 (2016). doi:10.1140/epjds/s13688-016-0093-1

work page doi:10.1140/epjds/s13688-016-0093-1 2016
[15]

Maria Teleki, Vedangi Bengali, Xiangjue Dong, Sai Tejas Janjur, Haoran Liu, Tian Liu, Cong Wang, Ting Liu, Yin Zhang, Frank Shipman, and James Caverlee. 2025. A Survey on LLMs for Story Generation. InFindings of the Association for Com- putational Linguistics: EMNLP 2025. Association for Computational Linguistics, Suzhou, China, 13954–13966. doi:10.18653/...

work page doi:10.18653/v1/2025.findings-emnlp.750 2025
[16]

2007.The Writer’s Journey: Mythic Structure for Writers(3rd ed.)

Christopher Vogler. 2007.The Writer’s Journey: Mythic Structure for Writers(3rd ed.). Michael Wiese Productions, Studio City, CA, USA

2007
[17]

Stephen G. Ware, R. Michael Young, Brent Harrison, and David L. Roberts. 2014. A Computational Model of Plan-Based Narrative Conflict at the Fabula Level. IEEE Transactions on Computational Intelligence and AI in Games6, 3 (Sept. 2014), 271–284. doi:10.1109/TCIAIG.2014.2326072

work page doi:10.1109/tciaig.2014.2326072 2014
[18]

Weiqi Wu, Hongqiu Wu, Lai Jiang, Xingyuan Liu, Hai Zhao, and Min Zhang. 2024. From Role-Play to Drama-Interaction: An LLM Solution. InFindings of the Asso- ciation for Computational Linguistics: ACL 2024. Association for Computational Linguistics, Bangkok, Thailand, 3271–3290. doi:10.18653/v1/2024.findings-acl.196

work page doi:10.18653/v1/2024.findings-acl.196 2024
[19]

Lvmin Zhang and Maneesh Agrawala. 2024. Transparent Image Layer Diffusion using Latent Transparency. arXiv:2402.17113 [cs.CV] https://arxiv.org/abs/2402. 17113

work page arXiv 2024
[20]

Drinkwater, and Mark O

Alexander Zook, Stephen Lee-Urban, Michael R. Drinkwater, and Mark O. Riedl
[21]

InProceedings of the Eighth Artificial Intelligence and Interactive Digi- tal Entertainment Conference (AIIDE ’12)

Skill-Based Mission Generation: A Data-Driven Temporal Player Modeling Approach. InProceedings of the Eighth Artificial Intelligence and Interactive Digi- tal Entertainment Conference (AIIDE ’12). Association for the Advancement of Artificial Intelligence, Stanford, CA, USA