Pith. sign in

REVIEW 3 major objections 4 minor 43 references

A new dataset of 44,226 in-game narrative choices shows frontier LLMs fail to beat a simple popularity baseline at predicting what a player picks next.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 09:25 UTC pith:LVU7EAS7

load-bearing objection Rushes is a genuinely useful preference dataset with a real but overinterpreted headline result; the data is worth a serious look, the RLHF inference needs an in-domain-trained model before it lands. the 3 major comments →

arxiv 2607.20767 v1 pith:LVU7EAS7 submitted 2026-07-22 cs.CL

Rushes: A Human Preference Dataset for Pluralistic Alignment

Michael Xu , Jorge Leandro , Sudha Rao , Weijia Xu , Nebojsa Jojic , Gabriel DesGarennes , Chris Quirk , Bill Dolan This is my paper
classification cs.CL
keywords pluralistic alignmentrevealed preferencesengagementinteractive narrativebranching storiespersonalized alignmentcollaborative filteringchoice prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Rushes is a dataset of revealed human engagement preferences collected from players of six AI-generated branching narrative games. The paper argues that the choices are structured and personalized: average choice entropy is below the uniform random baseline. The central result is an "Engagement Gap": when asked to predict a player's next choice from the story context and the player's previous decisions, state-of-the-art LLMs (including GPT-5) score about 34%—below the 36.4% of always picking the historically most popular option—while a classic collaborative-filtering model (SVD) reaches 37.7%. The authors read this as evidence that alignment methods which optimize for a single population-level preference are insufficient to capture the heterogeneous, context-dependent signals that drive felt engagement.

Core claim

The paper's core claim is that organic, sequential engagement choices carry a personalized signal that current large language models cannot extract, even when given the user's history. On event-level top-1 choice prediction across 8,293 held-out decisions, SVD matrix factorization achieves 37.73% accuracy, the popularity baseline 36.39%, GPT-5 with history 34.23%, SASRec 34.06%, a fine-tuned semantic classifier 30.0%, and random 25.41%. The dataset also shows that same-game history predicts far better than cross-game history (38.9% vs 29.1%), indicating that preferences are context-dependent. The paper calls the gap between personalized-history models and frontier LLMs the "Engagement Gap" a

What carries the argument

The central object is the Rushes dataset itself: a game interface in which players repeatedly choose one of four options at branching narrative points, generating time-ordered trajectories with per-user identifiers and full candidate sets. The benchmark task is event-level, text-based top-1 choice prediction from narrative context plus user history. The decisive mechanism is the comparison between SVD collaborative filtering—which learns per-user latent vectors from observed choices—and prompted LLMs given the user's textual history; the accuracy gap quantifies the proposed "Engagement Gap." The paper also uses choice entropy as evidence of non-random structure, and a user-stratified chronol

Load-bearing premise

The paper's broad conclusion depends on treating zero-shot, prompted LLMs with text-only history as a fair representative of alignment methods built on population-level objectives; no model in the paper is actually trained on Rushes with RLHF or DPO.

What would settle it

Fine-tune a model on the Rushes training split (e.g., DPO on the pairwise preference transforms) and evaluate on the same held-out split. If it matches or exceeds SVD's 37.7% and the popularity baseline's 36.4%, the claim that population-level objectives are insufficient would be undermined; if it still lags, the claim is supported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If Rushes is a valid diagnostic, current RLHF-style alignment that aggregates to a population-level reward may be structurally unable to produce engaging personalized interactive agents.
  • Models optimized for universal acceptability will tend to default to majority preferences, yielding generic rather than personally compelling choices.
  • Personalizing engagement in interactive narratives will require modeling per-user trajectories or otherwise capturing context-dependent history, not just scaling model size.
  • Benchmarks for pluralistic alignment should include revealed preferences from organic interaction, not only stated preferences or simulated personas.
  • The close parity between GPT-5 and SASRec suggests that sequential architecture alone does not close the gap; the paper implies the bottleneck is the objective, not the model class.

Where Pith is reading between the lines

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

  • The paper's critique of population-level objectives rests on a frozen, prompted LLM as the representative; a natural next step is to train an LLM or reward model on Rushes trajectories and test whether in-domain trained personalization recovers the gap.
  • The same-game vs cross-game history gap points to a strong role for narrative state, suggesting that user models may need to condition on the current story branch, not just a static user vector.
  • The 37.7% SVD accuracy is still far below ceiling, so the dataset likely supports more sophisticated personalization; the 'Engagement Gap' should be read as a lower-bound difficulty measure, not a proof that engagement is unpredictable.
  • The multimodal assets (image, video, audio) were generated but not used in the benchmark; a testable extension is whether providing them to an LLM changes prediction accuracy.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces Rushes, a dataset of 44,226 decision events from 8,167 users across six GPT-4o-generated branching narratives, with persistent user identifiers and full candidate sets. The authors propose event-level top-1 choice prediction as a benchmark for personalized engagement and pluralistic alignment. They report that a matrix factorization model (SVD) reaches 37.73% accuracy, the popularity baseline 36.39%, GPT-5 with user history 34.23%, SASRec 34.06%, and a fine-tuned DeBERTa-v3 semantic classifier 30.00%, against a random baseline of 25.41%. They interpret the gap between SVD and the LLMs as an 'Engagement Gap' and suggest that 'single, population-level objectives, like those used in modern RLHF, appear insufficient to capture heterogeneous, context-dependent engagement signals.' The paper also reports ablations on narrative depth, history source, user activity, and model scaling, and releases the dataset and generation code.

Significance. If the empirical claims hold, Rushes is a valuable resource for the community: it provides organic, revealed-preference data in a sequential, interactive setting, complements synthetic persona benchmarks such as PERSONA, and includes a clear evaluation protocol with user-stratified chronological splits, confidence intervals, and a detailed generation pipeline. The demonstration that a simple collaborative filter captures signal that prompted frontier LLMs do not is interesting and potentially useful for benchmarking personalized alignment. However, the paper's central interpretive conclusion—that the results indict population-level objectives used in modern RLHF—is not supported by the experiments as designed. The load-bearing comparison pits a frozen, prompted LLM against an in-domain-trained SVD model, and the statistical evidence for SVD's advantage over popularity is weaker than the prose suggests. These are fixable with additional experiments or by substantially qualifying the conclusion.

major comments (3)
  1. [Abstract; §4.1] The central claim that the Engagement Gap shows 'single, population-level objectives, like those used in modern RLHF, appear insufficient' rests on an untested equivalence. The LLM evaluations are zero-shot or prompted with text history; no reward model, DPO model, or RLHF-tuned LLM is trained on Rushes, even though §3.3.3 provides the pairwise transformation needed for such training. SVD is trained in-domain on Rushes user-option interactions, while the LLMs are frozen and prompted. The observed gap could disappear if an in-domain-trained reward model or RLHF-tuned model were evaluated. Please either add such an experiment or substantially weaken the conclusion to refer to frozen prompted LLMs, not the population-level objective class.
  2. [Table 4; §4.1] The evidence for personalized signal in SVD over popularity is statistically fragile. SVD's 37.73% [36.69–38.78] and popularity's 36.39% [35.36–37.43] have heavily overlapping 95% Wilson CIs, and the 1.34-point difference is reported without a paired significance test. Because decision events are clustered by user, game, and narrative path, independent-sample CIs likely overstate precision. A paired bootstrap over users or a per-user/per-item significance test is needed before claiming SVD 'captures measurable personalized signal.' Also clarify whether the popularity baseline is fit on the training split or the full dataset.
  3. [§4.2.1, Table 5] The depth analysis reports popularity accuracy increasing from 30.5% at depth 0 to 42.1% at depth 4, but the sample sizes vary widely (735 to 4,361) and no test for trend or interaction is provided. As a secondary ablation this is not blocking, but the current presentation overinterprets what may be sampling noise or differential attrition. If this depth trend is used to support the narrative-engagement interpretation, a per-depth significance test or at least explicit error bars on the comparisons is needed.
minor comments (4)
  1. [Figure 1 caption] Typo: 'theRushesdata' should read 'the Rushes data.' Similar missing spaces appear in several places (e.g., 'TextQuestsevaluates,' 'Rushescaptures').
  2. [§3.3.2, Figure 2] The entropy comparison (1.04 vs 1.39 nats) is reported without a significance test or confidence interval. Given the large N, the difference is likely real, but a test or interval would make the claim precise.
  3. [§3.3.1] The schema lists user_agent as logged metadata but the paper does not state whether this is retained in the released dataset or stripped to protect privacy. Please clarify.
  4. [Limitations] The limitations section discusses language and demographic skew but does not mention the frozen-prompted-LLM versus trained-model caveat that is central to the main conclusion. A sentence there would help readers calibrate the claim.

Circularity Check

0 steps flagged

No circularity: the headline results are direct held-out measurements; the RLHF inference is an untested interpretation, not a derived prediction.

full rationale

The paper's central quantitative claims (choice entropy, SVD accuracy, popularity baseline accuracy, GPT-4o/GPT-5 accuracy) are direct measurements on a user-stratified chronological split. SVD and popularity are fitted only on the training portion and are evaluated on the same held-out test split as the prompted LLMs; no equation in the paper reduces one measurement to another. The entropy comparison (1.04 nats vs. 1.39 nats uniform baseline) is a direct statistic, not a fitted parameter renamed as a prediction. The only self-referential element is that all narrative content and options were generated by GPT-4o, so the LLM baselines are evaluated on text produced by the same model family. This is a content-source confound, and the paper itself acknowledges it in the Limitations section ('This report relates to Rushes as implemented using GPT-4o. The results shown in the demonstration will differ if other LLMs are used.'). A confound of this kind may affect external validity, but it is not a circular derivation. The paper's inference that the observed gap suggests 'population-level objectives, such as those used in modern RLHF, may be insufficient' is an interpretation rather than a derivation: no RLHF-tuned model or reward model is trained on Rushes, so the inference is under-supported, but lack of support is not circularity. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The paper's quantitative claims are measurements, not derivations with fitted constants. One undisclosed configuration (SVD hyperparameters) affects the headline number. Five domain assumptions sit between the raw logs and the interpretive conclusions; none is machine-checked. No invented entities: the "Engagement Gap" and "pluralistic alignment" are labels for measured phenomena and a research program, not free-floating constructs with independent empirical handles.

free parameters (1)
  • SVD latent dimensionality / regularization = not reported
    The central personalized-signal claim (37.73%) depends on matrix-factorization hyperparameters; the paper specifies neither rank nor regularization, so the number cannot be reproduced or audited.
axioms (5)
  • domain assumption Uniform player distribution across branches (Appendix A.2, footnote: 'This assumption is used only to size the paraphrase budget; the actual distribution may be skewed.')
    Used to derive the paraphrase-count scaling K(d)≈√(P/b^(d+1)); harmless for the benchmark, but it is an assumption about traffic that A.2 acknowledges is likely false.
  • domain assumption LLM-based similarity checker correctly identifies redundant options (§3.1.1, Appendix A.8)
    The claim that users face "distinct and non-redundant alternatives" depends on this checker's judgment being accurate; failures would let near-duplicate options into candidate sets and inflate noise.
  • domain assumption Lower-than-uniform choice entropy indicates structured preferences (§3.3.2, Figure 2)
    Any distribution that is not perfectly uniform has lower entropy than the uniform baseline, so this statistic cannot distinguish genuine engagement structure from popularity skew or positional bias without a null model.
  • domain assumption Voluntary, unpaid in-game choices reveal engagement preferences (§3.3)
    The "revealed preference" interpretation assumes clicks reflect interest/engagement rather than interface artifacts (e.g., first-option bias, fatigue, random exploration); the paper provides no click-time or position-analysis check.
  • domain assumption GPT-4o-generated narratives are plausible engagement stimuli (§3.2.3)
    The authors state their goal was "plausible stimuli," not narrative quality; if the stories were confusing or repetitive, the measured "preferences" would partly reflect confusion.

pith-pipeline@v1.3.0-alltime-deepseek · 12752 in / 18038 out tokens · 156339 ms · 2026-08-01T09:25:13.281403+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Rushes: A Human Preference Dataset for Pluralistic Alignment." pith.science (2026). https://pith.science/paper/LVU7EAS7

@misc{pith2026260720767,
  author       = {Pith},
  title        = {Pith review of: Rushes: A Human Preference Dataset for Pluralistic Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LVU7EAS7}},
  note         = {Machine review of arXiv:2607.20767}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce Rushes, a dataset and benchmark for studying revealed human engagement preferences in interactive narrative environments. Rushes is collected through a game interface where users interact with AI-generated branching narratives and select one choice from a small, explicit candidate set at each decision point. Each interaction logs the full candidate set, the user's choice, and the evolving narrative context, yielding time-ordered trajectories with persistent user-level identifiers. Rushes contains 44,226 decision events from 8,167 unique users across six games, capturing sequential, personalized engagement behavior rather than static judgments. We show that user choices exhibit structured, non-random patterns, quantified by a low choice entropy relative to a uniform baseline. We position Rushes as a diagnostic benchmark for pluralistic alignment and demonstrate a robust Engagement Gap: state-of-the-art LLMs, including GPT-5, fail to outperform simple baselines. While classical Matrix Factorization (SVD) captures measurable personalized signal (37.7%), frontier LLMs (34.23%) struggle to even match the Popularity Baseline (36.4%) on event-level choice prediction. This gap suggests that single, population-level objectives, like those used in modern RLHF, appear insufficient to capture heterogeneous, context-dependent engagement signals. As a result, even highly capable models default to majority preferences rather than adapting to individual trajectories. We release Rushes to support research into pluralistic alignment and sequential decision-making in generative systems. The full code for the platform and dataset will be available here: https://github.com/microsoft/rushes

Figures

Figures reproduced from arXiv: 2607.20767 by Bill Dolan, Chris Quirk, Gabriel DesGarennes, Jorge Leandro, Michael Xu, Nebojsa Jojic, Sudha Rao, Weijia Xu.

Figure 1
Figure 1. Figure 1: Overview of the Rushes data collection framework. Left: The generation pipeline expands a user-provided synopsis into story segments and iteratively generates distinct options using similarity and quality checks. Center: The game interface presents the narrative and visual context, allowing players to select predefined choices or submit their own suggestions. Right: The data collection mechanism aggregates… view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of user vote entropy across de [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean semantic diversity, measured as pair [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

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

43 extracted references · 3 canonical work pages

  1. [1]

    Computer , year=

    Matrix Factorization Techniques for Recommender Systems , author=. Computer , year=

  2. [2]

    Self-Attentive Sequential Recommendation , booktitle =

    Kang, Wang-Cheng and McAuley, Julian , year =. Self-Attentive Sequential Recommendation , booktitle =

  3. [3]

    2025 , howpublished=

    GPT-5 System Card , author=. 2025 , howpublished=

    2025

  4. [4]

    Aidungeon

    AI Dungeon: Dragon Model Upgrade , author=. Aidungeon. io , year=

  5. [9]

    ICDM , year=

    Collaborative Filtering for Implicit Feedback Datasets , author=. ICDM , year=

  6. [10]

    ACM Computing Surveys , year=

    Sequence-aware Recommender Systems , author=. ACM Computing Surveys , year=

  7. [11]

    SIGIR , year=

    Deep Learning for Sequential Recommendation: Challenges and Methods , author=. SIGIR , year=

  8. [12]

    Advances in neural information processing systems , volume=

    Openassistant conversations-democratizing large language model alignment , author=. Advances in neural information processing systems , volume=

  9. [13]

    NeurIPS , year=

    Training Language Models to Follow Instructions with Human Feedback , author=. NeurIPS , year=

  10. [14]

    2024 , howpublished =

    GPT-4o System Card , author =. 2024 , howpublished =

    2024

  11. [16]

    2025 , howpublished =

    Demystifying Flux Architecture: FLUX.\!1 [schnell] , author =. 2025 , howpublished =

    2025

  12. [17]

    CoRR , year=

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback , author=. CoRR , year=

  13. [18]

    2024 , eprint=

    WHAT-IF: Exploring Branching Narratives by Meta-Prompting Large Language Models , author=. 2024 , eprint=

    2024

  14. [19]

    2025 , eprint=

    Fair-PP: A Synthetic Dataset for Aligning LLM with Personalized Preferences of Social Equity , author=. 2025 , eprint=

    2025

  15. [20]

    Drift: Decoding-time Personalized Alignments with Implicit User Preferences

    Kim, Minbeom and Lee, Kang-il and Joo, Seongho and Lee, Hwaran and Thonet, Thibaut and Jung, Kyomin. Drift: Decoding-time Personalized Alignments with Implicit User Preferences. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.324

    work page doi:10.18653/v1/2025.findings-emnlp.324 2025

  16. [21]

    Toward a Computational Framework of Suspense and Dramatic Arc

    O'Neill, Brian and Riedl, Mark. Toward a Computational Framework of Suspense and Dramatic Arc. Affective Computing and Intelligent Interaction. 2011

    2011

  17. [22]

    PERSONA : A Reproducible Testbed for Pluralistic Alignment

    Castricato, Louis and Lile, Nathan and Rafailov, Rafael and Fr. PERSONA : A Reproducible Testbed for Pluralistic Alignment. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  18. [23]

    2024 , eprint=

    Personalized Language Modeling from Personalized Human Feedback , author=. 2024 , eprint=

    2024

  19. [24]

    2024 , eprint=

    Personalizing Reinforcement Learning from Human Feedback with Variational Preference Learning , author=. 2024 , eprint=

    2024

  20. [25]

    Electronics , VOLUME =

    Yu, Yaojiong and Corino, Gianni and Phillips, Mike , TITLE =. Electronics , VOLUME =. 2025 , NUMBER =

    2025

  21. [26]

    Choice-75: A Dataset on Decision Branching in Script Learning

    Hou, Zhaoyi and Zhang, Li and Callison-Burch, Chris. Choice-75: A Dataset on Decision Branching in Script Learning. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). 2024

    2024

  22. [29]

    2025 , eprint=

    LiteraryTaste: A Preference Dataset for Creative Writing Personalization , author=. 2025 , eprint=

    2025

  23. [30]

    2025 , eprint=

    Beyond Correctness: Evaluating Subjective Writing Preferences Across Cultures , author=. 2025 , eprint=

    2025

  24. [32]

    2025 , eprint=

    Operationalizing Pluralistic Values in Large Language Model Alignment Reveals Trade-offs in Safety, Inclusivity, and Model Behavior , author=. 2025 , eprint=

    2025

  25. [33]

    Dalia Ali, Dora Zhao, Allison Koenecke, and Orestis Papakyriakopoulos. 2025. http://arxiv.org/abs/2511.14476 Operationalizing pluralistic values in large language model alignment reveals trade-offs in safety, inclusivity, and model behavior

    arXiv 2025

  26. [34]

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. CoRR

    2022

  27. [35]

    Louis Castricato, Spencer Frazier, Jonathan Balloch, and Mark Riedl. 2021. https://doi.org/10.18653/v1/2021.nuse-1.9 Fabula entropy indexing: Objective measures of story coherence . In Proceedings of the Third Workshop on Narrative Understanding, pages 84--94, Virtual. Association for Computational Linguistics

    work page doi:10.18653/v1/2021.nuse-1.9 2021

  28. [36]

    Louis Castricato, Nathan Lile, Rafael Rafailov, Jan-Philipp Fr \"a nken, and Chelsea Finn. 2025. https://aclanthology.org/2025.coling-main.752/ PERSONA : A reproducible testbed for pluralistic alignment . In Proceedings of the 31st International Conference on Computational Linguistics, pages 11348--11368, Abu Dhabi, UAE. Association for Computational Linguistics

    2025

  29. [37]

    Halperin, Yuwen Lu, and Max Kreminski

    John Joon Young Chung, Vishakh Padmakumar, Melissa Roemmele, Yi Wang, Yuqian Sun, Tiffany Wang, Shm Garanganao Almeda, Brett A. Halperin, Yuwen Lu, and Max Kreminski. 2025. http://arxiv.org/abs/2511.09310 Literarytaste: A preference dataset for creative writing personalization

    arXiv 2025

  30. [38]

    Angela Fan, Mike Lewis, and Yann Dauphin. 2018. https://doi.org/10.18653/v1/P18-1082 Hierarchical neural story generation . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 889--898, Melbourne, Australia. Association for Computational Linguistics

    doi:10.18653/v1/p18-1082 2018

  31. [39]

    Parsa Ghaffari and Chris Hokamp. 2025. http://arxiv.org/abs/2504.02426 Narrative studio: Visual narrative exploration using LLMs and Monte Carlo Tree Search

    Pith/arXiv arXiv 2025

  32. [40]

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. 2025. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103

    Pith/arXiv arXiv 2025

  33. [41]

    Martin, and Chris Callison-Burch

    Runsheng "Anson" Huang, Lara J. Martin, and Chris Callison-Burch. 2024. http://arxiv.org/abs/2412.10582 What-if: Exploring branching narratives by meta-prompting large language models

    arXiv 2024

  34. [42]

    Wang-Cheng Kang and Julian McAuley. 2018. https://doi.org/10.1109/ICDM.2018.00035 Self-attentive sequential recommendation . In 2018 IEEE International Conference on Data Mining (ICDM), pages 197--206

    arXiv 2018

  35. [43]

    OpenAI . 2024. Gpt-4o system card. arXiv preprint, https://arxiv.org/abs/2410.21276. Accessed 2025-09-22

    Pith/arXiv arXiv 2024

  36. [44]

    OpenAI. 2025. Gpt-5 system card. https://openai.com/index/gpt-5-system-card. Accessed: 2025-09-22

    2025

  37. [45]

    Long Ouyang et al. 2022. Training language models to follow instructions with human feedback. In NeurIPS

    2022

  38. [46]

    Long Phan, Mantas Mazeika, Andy Zou, and Dan Hendrycks. 2025. http://arxiv.org/abs/2507.23701 Textquests: How good are LLMs at text-based video games?

    Pith/arXiv arXiv 2025

  39. [47]

    Md Awsafur Rahman, Adam Gabrys, Doug Kang, Jingjing Sun, Tian Tan, and Ashwin Chandramouli. 2025. http://arxiv.org/abs/2512.13077 Likebench: Evaluating subjective likability in LLMs for personalization

    arXiv 2025

  40. [48]

    Riedl and Vadim Bulitko

    Mark O. Riedl and Vadim Bulitko. 2013. https://doi.org/10.1609/aimag.v34i1.2449 Interactive narrative: An intelligent systems approach . AI Magazine, 34(1):67--77

    doi:10.1609/aimag.v34i1.2449 2013

  41. [49]

    Nick Walton. 2019. https://aidungeon.io/ Ai dungeon: Dragon model upgrade . Aidungeon. io

    2019

  42. [50]

    Shuangshuang Ying, Yunwen Li, Xingwei Qu, Xin Li, Sheng Jin, Minghao Liu, Zhoufutu Wen, Xeron Du, Tianyu Zheng, Yichi Zhang, Letian Ni, Yuyang Cheng, Qiguang Chen, Jingzhe Ding, Shengda Long, Wangchunshu Zhou, Jiazhan Feng, Wanjun Zhong, Libo Qin, Ge Zhang, Wenhao Huang, Wanxiang Che, and Chenghua Lin. 2025. http://arxiv.org/abs/2510.14616 Beyond correctn...

    arXiv 2025

  43. [51]

    Hong Yu and Mark Riedl. 2013. https://doi.org/10.1609/aiide.v9i1.12665 Data-driven personalized drama management . Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 9(1):191--197

    work page doi:10.1609/aiide.v9i1.12665 2013