Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

All Stories Are One Story: Emotional Arc Guided Procedural Game Level Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Story arcs guide procedural game levels to higher enjoyment by serving as a generative constraint on narrative and difficulty.

desk verdict A genuinely useful LLM+PCG systems paper whose main enjoyment result is undermined by an internal difficulty-control contradiction; worth serious review, but the causal claim needs fixing. read the letter →

arxiv 2508.02132 v1 pith:Z43ZUP7N submitted 2025-08-04 cs.AI

Yunge Wen , Chenliang Huang , Hangyu Zhou , Zhuo Zeng , Chun Ming Louis Po , Julian Togelius , Timothy Merino , Sam Earle This is my paper
classification cs.AI
keywords emotionalarcsproceduralcontentgenerationlargelanguagemodelsbranchingnarrativeplayerexperiencesentimentanalysisgamelevelinteractivestorytelling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that emotional arcs—the rise-and-fall patterns of affective tone that recur in stories—can be used as a working generative constraint for procedural game level generation, not just as a descriptive theory. It builds a pipeline that takes a short user prompt and an arc type (Rise, Fall, or a canonical combination such as Cinderella) and produces a branching story graph whose nodes are labeled Rise or Fall, with each node automatically populated with characters, items, and difficulty-adjusted enemy stats. In a prototype action RPG, the arc-guided episodes scored significantly higher on enjoyment (p=0.016, r=0.789), and 13 of 16 players correctly identified which episode contained an arc. The authors argue this demonstrates that emotional structure can serve as a scaffold for narrative coherence and emotional impact in AI-driven game design.

What carries the argument

The load-bearing object is the pair of primitive emotional labels, Rise and Fall, taken from a six-arc taxonomy of stories. Each node of a directed acyclic story graph receives one of these labels, and the label drives two coupled processes: a 'mind reset' prompt that rewrites the node's prose with an uplifting or somber tone, and automatic configuration of enemy statistics so Fall nodes are harder and more hostile while Rise nodes are friendlier and easier. The graph's edges carry playable triggers ('defeat an enemy', 'talk to an NPC') that map onto dungeon rooms in a Unity prototype, and the whole pipeline is structured as an AI chain that separates story-graph construction, semantic refinement, and entity generation so designers can revise nodes before levels are built.

What would settle it

Generate many stories for each of the six arcs using the exact prompt strings and test whether a strict monotonicity check on per-node valence passes for Rise and Fall; the paper's own data already suggest it would not. A sharper falsifier is to ask blind players to sort monotonic-arc stories into the intended arc order at better-than-chance rates, which would isolate whether the label itself drives the enjoyment effect rather than any generic improvement in prose quality.

Watch

Extended reading notes

Core claim

The central claim is that embedding a canonical emotional arc into the generation process—by assigning each story node a Rise or Fall label, then using those labels to set the node's textual tone and to scale entity difficulty—produces procedural game narratives that players enjoy more and perceive as more emotionally coherent than affect-neutral baselines. The quantitative anchor is the Enjoyment result: a one-sided Wilcoxon signed-rank test gave p=0.016 with a large effect size (r=0.789), and the bootstrap confidence intervals for Enjoyment did not overlap between conditions. The paper also reports that 13 out of 16 participants (81.25%) correctly identified the arc condition, and that valence trajectories from an independent sentiment classifier matched the intended shapes for non-monotonic arcs such as Rise–Fall–Rise, which the authors take as external validation that the intended narrative structure survives in the generated emotional flow.

Load-bearing premise

The whole pipeline rests on the assumption that the Rise and Fall labels, injected as short tone-setting prompts, reliably produce distinguishable emotional trajectories in the generated text—yet the paper's own sentiment analysis shows that the simplest arcs, steady Rise and steady Fall, come out flat or moderately positive rather than following their intended directions.

Editorial extensions

If this is right

  • Game developers can use a high-level emotional arc as a generative constraint that improves player enjoyment without hand-authoring every branch of a story.
  • LLM-based procedural generation can maintain global emotional coherence—not just local textual fluency—when arc labels are injected as explicit prompts.
  • The same arc label can synchronize narrative tone and gameplay difficulty, since Fall nodes empirically produced stronger enemies and Rise nodes friendlier ones.
  • Non-monotonic arcs such as Rise–Fall–Rise are reproducible by sentiment analysis, suggesting they are reliable tools for controlled emotional pacing.
  • The 81.25% identification rate implies arcs are perceptible to players, making them a usable design signal rather than an invisible theoretical property.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The framework's reliance on text-heavy narrative likely explains why some players preferred the baseline; a testable extension is to express arc cues through non-textual systems (enemy behavior, lighting, music) and measure whether perceptibility rises while 'too much text' complaints fall.
  • The monotonic arcs' failure to show directional trends suggests the two prompt strings alone are too weak; an explicit next step is valence supervision per node, such as conditioning generation on target positivity scores.
  • The positive-valence bias across all arcs hints that both the story generator and the sentiment classifier skew positive; calibrating the classifier on neutral text would clarify whether the flat Rise/Fall trajectories reflect generation or measurement.
  • Because the evaluation is confined to one ARPG template, the framework's promise of genre transfer remains untested; roguelikes and visual novels would be direct testbeds for the same node-and-edge pipeline.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper presents an LLM-based pipeline for procedural game narrative generation in which branching story graphs are labeled with Rise/Fall emotional arcs drawn from Reagan et al. (2016). Each story node is automatically populated with narrative text, characters, items, and gameplay attributes, with enemy difficulty purportedly aligned to the emotional trajectory. The system is instantiated in an ARPG prototype with generated pixel-art sprites and Unity levels. Evaluation consists of a within-subject user study (n=16) comparing a Baseline episode with an Emotional Arc episode on enjoyment, relevance, and difficulty ratings; a binary arc-identification task; semi-structured interviews; and a post-hoc GoEmotions sentiment analysis of generated arcs. The main reported quantitative result is a significant enjoyability advantage for the Emotional Arc condition (Wilcoxon p=0.016, r=0.789), and 13/16 participants correctly identified the arc condition. Sentiment trajectories show alignment for non-monotonic arcs but not for monotonic Rise/Fall arcs.

Significance. If the central claim is sustained, the framework would be a practical contribution: it gives designers a high-level emotional constraint that can be operationalized by LLMs into playable, branching game content, and it tests the result with both human perception and an independent sentiment classifier. Strengths of the submission include the end-to-end system implementation, the use of an outside emotion classifier rather than the generating LLM itself, and the transparent reporting of qualitative user feedback. However, the significance is currently limited by the small sample, the exploratory design, and an unresolved confound between narrative arc and gameplay difficulty. The paper's value therefore depends on whether the authors can disentangle the emotional arc from difficulty pacing in the only statistically significant result.

major comments (3)
  1. [Evaluation (User Study) vs. Method (Entity Generation) and Case Study] The causal attribution of the Enjoyment result (Table 2, p=0.016, r=0.789) to emotional arc integration is undermined by an internal contradiction. The Method states that entity statistics are 'automatically configured based on the emotional arc'; the Case Study reports that enemies in Fall nodes were more likely to have higher attack, health, and ranged capabilities; the Evaluation claims that 'gameplay mechanics, and difficulty parameters were held constant across conditions'; and the Limitations concedes that the Baseline condition 'did not include consistent difficulty modulation' and that ranged-attack enemies may have confounded difficulty comparisons. No per-condition entity statistics are reported. As written, the significant enjoyment difference could reflect difficulty pacing or enemy composition rather than the emotional narrative structure, so the only significant quantitative outcome for RQ1 cannot be causally interpreted. The authors need to report the actual entity-stat distributions per condition and, if those are not available, temper the RQ1 conclusion or provide an additional controlled experiment.
  2. [Results (Sentiment Alignment)] The system-level validation fails for the two primitive labels on which the whole pipeline is built. The paper states that 'monotonic arc types (steady Rise, steady Fall) do not exhibit strong directional trends' and remain flat or moderately positive. Since every canonical arc is composed of Rise and Fall labels and the mind reset prompts condition text on these two labels, the sentiment analysis shows that the basic building blocks are not reliably conveyed to the text. The RQ2 claim that emotional arcs are perceptible to computational emotion models is therefore only supported for non-monotonic composites, not for Rise and Fall themselves. This weakens the central claim that emotional arcs are a practical generative constraint and should be acknowledged explicitly in the abstract and discussion, not only in the Limitations section.
  3. [Abstract, Discussion, and Conclusion] The abstract and conclusion claim that emotional arc integration 'significantly enhances engagement, narrative coherence, and emotional impact.' The only statistically significant quantitative user-rating is Enjoyment; Relevance is not significant (p=0.086), Difficulty is not significant (p=0.573), and no direct Likert or otherwise quantified measure of 'narrative coherence' or 'emotional impact' is reported. The qualitative interview themes are suggestive but cannot carry the weight of a significance claim in the abstract. The claims should be scaled to what the data actually support: a significant enjoyment difference whose attribution is currently confounded, plus qualitative and sentiment-based evidence of perceptibility for non-monotonic arcs.
minor comments (4)
  1. [Evaluation (User-Rated Scales)] In the text, bootstrapping is described as using 1,000 samples, but Table 1's caption says 10,000 samples; the discrepancy should be reconciled.
  2. [Results (Figure 4)] Figure 4 would be easier to interpret if the y-axis were explicitly labeled with the valence scale (e.g., the computed V values), since the reader currently has to infer the range and units from the text.
  3. [Method (Story Generation)] Reproducibility would be improved by providing the exact StoryChainRevision prompt templates or an appendix with the full prompts, rather than only the mind reset strings.
  4. [References] The Todd et al. reference is incomplete: 'In Proceedings of the 2023 ACM Conference' omits the conference name and page range.

Circularity Check

0 steps flagged · score 2.0 of 10

No central circularity: emotional arcs are inputs, and the main quantitative support is an external user study; only minor self-citations and a manipulation-check sentiment analysis keep the score above zero.

full rationale

The paper's derivation chain does not reduce to its inputs. Emotional arc labels (Rise/Fall) are user-provided inputs to the LLM pipeline, not parameters fitted to the evaluation data. The Enjoyment result (p = 0.016, r = 0.789) comes from a between-condition user study with an independent behavioral rating, and the 13/16 arc-identification result is an independent perceptual judgment. The sentiment-alignment analysis in 'Sentiment Alignment' and Figure 4 is best read as a manipulation check: the LLM was explicitly prompted with mind reset strings such as 'Embrace an uplifting and hopeful tone' and 'Adopt a more somber and challenging tone', so the measured valence aligning with those instructions largely verifies prompt adherence rather than testing an independent prediction. That is weak validation, but it is not a fitted input renamed as a prediction and does not bear the central claim. Self-citations (Todd et al. 2023; Liapis et al. 2013; Alvarez et al. 2022) appear only as background in Related Works and are not load-bearing. The Limitations section candidly concedes that the Baseline condition lacked consistent difficulty modulation and that ranged-attack enemies may have confounded difficulty comparisons; this is an internal-validity concern for attributing the Enjoyment effect to emotional arcs, not a circular derivation. No equation or parameter is defined in terms of the target result, no uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed. Accordingly, the circularity score is 2 rather than 0 only because of the minor non-load-bearing self-citations and the manipulation-check nature of the sentiment validation.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central empirical result rests on small samples and subjective ratings; the sentiment validation is only weakly independent of the generation prompts. No free parameters were fitted to produce the enjoyment result, but the valence analysis depends on hand-set thresholds.

free parameters (1)
  • Simplified GoEmotions valence mapping and label retention threshold = probability threshold 0.1; valence +1/0/-1
    Labels with probability >=0.1 are retained and mapped to valence, a hand-chosen setting that directly shapes the sentiment trajectories in Figure 4 and the claimed arc alignment.
assumptions (4)
  • domain assumption The six emotional arcs of Reagan et al. (2016) are a universal and transferable narrative structure.
    The method takes these arcs as the generative backbone without testing their applicability to interactive game play; the monotonic arcs already fail to reproduce their shapes in the sentiment analysis.
  • ad hoc to paper The mind reset prompt strings reliably set the emotional tone of the generated storyline text.
    The entire emotional realization depends on these prompt phrases and the LLM's compliance; no verification that the model follows them beyond the mixed sentiment trajectories.
  • domain assumption GoEmotions labels and the +1/0/-1 valence mapping measure the emotional experience of players.
    The classifier captures text sentiment, not player affect; using it as external validation assumes the two align.
  • ad hoc to paper Entity statistics (health, attack, ranged) generated from emotional labels track difficulty without confounding the user study.
    The paper both asserts difficulty was held constant and describes emotional arc difficulty modulation, so the assumption underlying the comparison is unclear.

how reviews work

0 comments
Cite this review

Pith. "Pith review of All Stories Are One Story: Emotional Arc Guided Procedural Game Level Generation." pith.science (2026). https://pith.science/paper/Z43ZUP7N

@misc{pith2026250802132,
  author       = {Pith},
  title        = {Pith review of: All Stories Are One Story: Emotional Arc Guided Procedural Game Level Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z43ZUP7N}},
  note         = {Machine review of arXiv:2508.02132}
}
read the original abstract

The emotional arc is a universal narrative structure underlying stories across cultures and media -- an idea central to structuralist narratology, often encapsulated in the phrase "all stories are one story." We present a framework for procedural game narrative generation that incorporates emotional arcs as a structural backbone for both story progression and gameplay dynamics. Leveraging established narratological theories and large-scale empirical analyses, we focus on two core emotional patterns -- Rise and Fall -- to guide the generation of branching story graphs. Each story node is automatically populated with characters, items, and gameplay-relevant attributes (e.g., health, attack), with difficulty adjusted according to the emotional trajectory. Implemented in a prototype action role-playing game (ARPG), our system demonstrates how emotional arcs can be operationalized using large language models (LLMs) and adaptive entity generation. Evaluation through player ratings, interviews, and sentiment analysis shows that emotional arc integration significantly enhances engagement, narrative coherence, and emotional impact. These results highlight the potential of emotionally structured procedural generation for advancing interactive storytelling for games.

Figures

Figures reproduced from arXiv: 2508.02132 by the authors.

Figure 1
Figure 1. End-to-end framework for emotional arc–guided game generation. The pipeline takes a user-provided narrative prompt [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. End-to-end narrative generation and gameplay interface. (A) Users input a story prompt to generate an emotional arc. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Boxplot comparison of user ratings across three [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Emotion valence trajectories for six canonical emotional arcs across story levels (1–7). Each subplot corresponds [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. High-quality generation of dynamic game content via small language models: A proof of concept

    cs.AI 2026-01 conditional novelty 5.0 of 10

    Proof-of-concept shows fine-tuned small language models achieve adequate quality for real-time game content generation in a scoped RPG loop via retry-until-success and LLM-as-judge evaluation.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Alvarez, A.; Font, J.; and Togelius, J. 2022. Story Designer: Towards a Mixed-Initiative Tool to Create Narrative Structures. In Proceedings of the 17th International Conference on the Foundations of Digital Games, 1--9

    work page 2022

  4. [4]

    J.; and Riedl, M

    Ammanabrolu, P.; Tien, E.; Cheung, W.; Luo, Z.; Ma, W.; Martin, L. J.; and Riedl, M. O. 2020. Story Realization: Expanding Plot Events into Sentences. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, 7375--7382

    work page 2020

  5. [5]

    T.; Li, Y.; Lundberg, S.; Nori, H.; Palangi, H.; Ribeiro, M

    Bubeck, S.; Chandrasekaran, V.; Eldan, R.; Gehrke, J.; Horvitz, E.; Kamar, E.; Lee, P.; Lee, Y. T.; Li, Y.; Lundberg, S.; Nori, H.; Palangi, H.; Ribeiro, M. T.; and Zhang, Y. 2023. Sparks of Artificial General Intelligence: Early Experiments with GPT-4. arXiv preprint arXiv:2303.12712

    arXiv 2023

  6. [6]

    Buongiorno, S.; Klinkert, L.; Zhaung, Z.; Chawla, T.; and Clark, C. 2024. PANGeA : Procedural Artificial Narrative Using Generative AI for Turn-Based, Role-Playing Video Games. In Proceedings of the Twentieth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, AIIDE '24. AAAI Press

    work page 2024

  7. [7]

    Calderwood, A.; Wardrip-Fruin, N.; and Mateas, M. 2022. Spinning Coherent Interactive Fiction through Foundation Model Prompts. In Proceedings of the 13th International Conference on Computational Creativity, 44--53. Bozen-Bolzano, Italy: Association for Computational Creativity

    work page 2022

  8. [8]

    Campbell, J. 2008. The Hero with a Thousand Faces. Joseph Campbell Foundation; New World Library, 3rd edition. Originally published in 1949

    work page 2008

Show all 25 references
  1. [9]

    Grabska-Gradzinska, I.; Nowak, L.; Palacz, W.; and Grabska, E. 2021. Application of Graphs for Story Generation in Video Games. In Proceedings of the 2021 Australasian Computer Science Week Multiconference, 1--6

    2021

  2. [10]

    Kumaran, V.; Rowe, J.; and Lester, J. 2024. NarrativeGenie: Generating Narrative Beats and Dynamic Storytelling with Large Language Models. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE), volume 20, 76--86

    2024

  3. [11]

    Leandro, J.; Rao, S.; Xu, M.; Xu, W.; Jojic, N.; Brockett, C.; and Dolan, B. 2024. GENEVA: GENErating and Visualizing Branching Narratives Using LLMs. In IEEE Conference on Games 2024

    2024

  4. [12]

    Liapis, A.; Yannakakis, G.; and Togelius, J. 2013. Designer Modeling for Personalized Game Content Creation Tools. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, volume 9, 11--16

    2013

  5. [13]

    U.; James, S.; and Togelius, J

    Nasir, M. U.; James, S.; and Togelius, J. 2024. Word2World: Generating Stories and Worlds through Large Language Models. arXiv preprint

    2024

  6. [14]

    Naughty Dog . 2013. The Last of Us. PlayStation 3. Published by Sony Computer Entertainment

    2013

  7. [15]

    Polti, G. 1921. The Thirty-Six Dramatic Situations. Boston, MA: The Editor Company

    1921

  8. [16]

    Propp, V. 1928. Morphology of the Folktale, volume 9 of Publications of the American Folklore Society. Austin, TX: University of Texas Press, second edition. ISBN 9780292792494. Originally published in Russian in 1928

    1928

  9. [17]

    J.; Mitchell, L.; Kiley, D.; Danforth, C

    Reagan, A. J.; Mitchell, L.; Kiley, D.; Danforth, C. M.; and Dodds, P. S. 2016. The emotional arcs of stories are dominated by six basic shapes. EPJ Data Science, 5(31)

    2016

  10. [18]

    O.; and Bulitko, V

    Riedl, M. O.; and Bulitko, V. 2013. Interactive Narrative: An Intelligent Systems Approach. AI Magazine, 34(1): 67--77

    2013

  11. [19]

    O.; and Young, R

    Riedl, M. O.; and Young, R. M. 2010. Narrative Planning: Balancing Plot and Character. Journal of Artificial Intelligence Research, 39: 217--268

    2010

  12. [20]

    Sudhakaran, S.; Gonz \'a lez-Duque, M.; Freiberger, M.; Glanois, C.; Najarro, E.; and Risi, S. 2023. Mariogpt: Open-ended text2level generation through large language models. Advances in Neural Information Processing Systems, 36: 54213--54227

    2023

  13. [21]

    Supermassive Games . 2015. Until Dawn. PlayStation 4. Published by Sony Computer Entertainment

    2015

  14. [22]

    U.; Green, M

    Todd, G.; Earle, S.; Nasir, M. U.; Green, M. C.; and Togelius, J. 2023. Level Generation Through Large Language Models. In Proceedings of the 2023 ACM Conference. New York, NY, USA: Association for Computing Machinery. ISBN 9781450398558

    2023

  15. [23]

    G.; Markowitz, D.; Boyd, A

    Ware, S. G.; Markowitz, D.; Boyd, A. M.; and Cardona-Rivera, R. E. 2022. Multiagent Narrative Experience Management as Story Graph Pruning. In Proceedings of the Eighteenth AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE), 108--114

    2022

  16. [24]

    Wu, T.; Terry, M.; and Cai, C. J. 2022. AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts. In CHI Conference on Human Factors in Computing Systems (CHI '22), 1--22. New York, NY, USA. ISBN 978-1-4503-9157-3/22/04

    2022

  17. [25]

    Zhang, L.; and Agrawala, M. 2024. Transparent Image Layer Diffusion using Latent Transparency. arXiv:2402.17113

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.