arxiv: 2604.11655 · v1 · submitted 2026-04-13 · 💻 cs.CL · cs.AI· cs.MA

Recognition: unknown

RPA-Check: A Multi-Stage Automated Framework for Evaluating Dynamic LLM-based Role-Playing Agents

Riccardo Rosati , Edoardo Colucci , Massimiliano Bolognini , Adriano Mancini , Paolo Sernani

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:13 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.MA

keywords role-playing agentsLLM evaluationautomated frameworkboolean indicatorsmodel size trade-offsprocedural consistencylegal scenarios

0 comments

The pith

A four-step pipeline converts role-playing criteria into boolean checklists scored by an LLM judge to evaluate agent fidelity.

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

The paper presents RPA-Check as a method to assess how well LLM-based agents stick to assigned roles in open-ended interactions where usual metrics fall short. It starts by defining behavioral dimensions, expands them into specific yes-no indicators, filters those for independence and clarity, then uses chain-of-thought prompting in another model to produce scores. This approach is tested on a legal training game with multiple quantized models across five scenarios. The results indicate that the system can surface trade-offs, including cases where smaller instruction-tuned models maintain better procedural consistency than larger ones.

Core claim

RPA-Check evaluates LLM role-playing agents by first defining high-level criteria, augmenting them into granular boolean indicators, applying semantic filtering to ensure no redundancy and full isolation, and then running an LLM-as-a-Judge step with chain-of-thought verification to score fidelity. When applied to several quantized models in a forensic training environment across five legal scenarios, the framework identifies an inverse relationship between parametric scale and consistency, with 8-9B models showing stronger performance than larger architectures affected by user-alignment bias.

What carries the argument

The RPA-Check pipeline that turns qualitative role requirements into non-redundant boolean indicators for automated LLM-based scoring.

If this is right

The method can expose specific trade-offs between model size, reasoning depth, and long-term stability in agent behavior.
Smaller instruction-tuned models can deliver higher procedural consistency than larger ones in constraints-heavy settings.
It supplies a reproducible metric for comparing generative agents in specialized domains.
Findings point to the value of targeted instruction tuning over raw parameter count for maintaining role fidelity.

Where Pith is reading between the lines

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

Teams building interactive agents may achieve better reliability by selecting mid-sized tuned models rather than scaling up.
The checklist approach could transfer to evaluating consistency in other long-horizon domains such as educational simulations.
Combining the automated scores with occasional human review might further reduce judge bias in high-stakes uses.

Load-bearing premise

That an LLM judge using chain-of-thought reasoning can produce objective scores for role adherence without its own biases or missed failure modes.

What would settle it

A new set of role-playing tests in which larger models receive equal or higher consistency scores than smaller ones while avoiding sycophancy, or where the boolean indicators fail to flag a major observed breakdown in narrative stability.

Figures

Figures reproduced from arXiv: 2604.11655 by Adriano Mancini, Edoardo Colucci, Massimiliano Bolognini, Paolo Sernani, Riccardo Rosati.

**Figure 1.** Figure 1: Architecture of the RPA-Check Framework. The automated evaluation pipeline transforms qualitative behavioral requirements of Role-Playing agents into quantitative metrics through a four-stage process: Dimensions Definition establishes behavioral, procedural, and linguistic criteria; Augmentation expands these dimensions into granular Boolean checklists via an LLM generator; Semantic Filtering ensures check… view at source ↗

**Figure 2.** Figure 2: Screens from LLM Court: (a) the main game menu and (b) the case generation module. During the simulation, the game transitions to the courtroom, where the judge RPA (c) introduces the case, and the player (d) interacts with the RPAs through text-based input. click new game Loading of LLM architecture Show preview of generated case Launching the game and the main interface Random data generation using large… view at source ↗

**Figure 3.** Figure 3: The player’s flow in LLM Court. After launching the game interface, starting a new game loads the RPAs LLM. Upon selecting a “new case,” the Procedural Case Generation module powered by GPT-5 generates and displays a case description. The player may either save the generated case to begin the game or discard it and request a new one. This module deserializes the output into run-time objects, populating the… view at source ↗

**Figure 4.** Figure 4: Schematic of the LLM Court simulation flow. The diagram illustrates the multi-phase execution logic managed by the FSM. The pools delineate the operational boundaries between the system-level procedural logic, the AI, i.e., the LLM-driven RPAs, and the interactive human player input (Defense). Process Logic is divided into three distinct chronological stages: the Introduction Phase (fixed sequential setup)… view at source ↗

read the original abstract

The rapid adoption of Large Language Models (LLMs) in interactive systems has enabled the creation of dynamic, open-ended Role-Playing Agents (RPAs). However, evaluating these agents remains a significant challenge, as standard NLP metrics fail to capture the nuances of role adherence, logical consistency, and long-term narrative stability. This paper introduces RPA-Check, a multi-stage automated evaluation framework designed to objectively assess the performance of LLM-based RPAs in complex, constraints-heavy environments. Our methodology is based on a four-step pipeline: (1) Dimension Definition, establishing high-level qualitative behavioral criteria; (2) Augmentation, where these requirements are expanded into granular boolean checklist indicators; (3) Semantic Filtering, to ensure indicator objectivity, no redundancy and agent isolation; and (4) LLM-as-a-Judge Evaluation, which employs chain-of-thought verification to score agent fidelity. We validate this framework by applying it to LLM Court, a serious game for forensic training involving several quantized local models. Experimental results across five distinct legal scenarios demonstrate the framework's ability to identify subtle trade-offs between model size, reasoning depth, and operational stability. Notably, the findings reveal an inverse relationship between parametric scale and procedural consistency, showing that smaller, adequately instruction-tuned models (8-9B) can outperform larger architectures prone to user-alignment bias or sycophancy. RPA-Check thus provides a standardized and reproducible metric for future research in generative agent evaluation within specialized domains.

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

RPA-Check lays out a four-stage checklist pipeline for judging role-playing agents in tight domains like legal training, but the headline claim that smaller models beat larger ones on consistency rests on unvalidated LLM judges.

read the letter

The paper's main contribution is a concrete four-step process for evaluating dynamic role-playing agents: define behavioral dimensions, expand them into boolean indicators, filter for non-redundancy and isolation, then score with chain-of-thought LLM judgment. This targets a clear gap where standard metrics miss role adherence and long-term stability in constrained settings such as the LLM Court forensic game. The pipeline itself is a reasonable attempt to make evaluation more systematic and reproducible than ad-hoc prompting.

Referee Report

2 major / 1 minor

Summary. The paper introduces RPA-Check, a four-stage automated evaluation framework for LLM-based role-playing agents consisting of (1) dimension definition, (2) augmentation to granular boolean checklist indicators, (3) semantic filtering for objectivity, non-redundancy and agent isolation, and (4) chain-of-thought LLM-as-a-Judge scoring of agent fidelity. The framework is validated by application to the LLM Court serious game across five legal scenarios with quantized local models, with the central empirical claim being an inverse relationship between parametric scale and procedural consistency, such that smaller instruction-tuned models (8-9B) outperform larger architectures due to reduced user-alignment bias or sycophancy.

Significance. If the reported trade-offs prove robust under independent validation, the framework could supply a reproducible, domain-specific metric for assessing role adherence and narrative stability in interactive agents, particularly in high-stakes settings such as forensic training. The scale-consistency observation, if not an artifact of the evaluation pipeline, would have practical implications for model selection in constrained role-play applications.

major comments (2)

[Methodology, step 4] LLM-as-a-Judge Evaluation (step 4): The headline finding of an inverse relationship between model scale and consistency rests entirely on scores produced by an LLM judge with chain-of-thought. No details are supplied on the identity or size of the judge model, any ablation of alternative judges, or any correlation of the resulting scores against human expert ratings on the same boolean indicators. Without such grounding, systematic preferences in the judge could mechanically generate the reported sycophancy penalty on larger models.
[Experiments and Results] Experimental results across five scenarios: The manuscript asserts that the results demonstrate subtle trade-offs and the inverse scale-consistency relationship, yet the provided text contains no tables of per-model scores, error bars, statistical significance tests, exclusion criteria, or raw indicator-level data. This absence prevents assessment of whether the 8-9B advantage is reliable or driven by a small number of scenarios.

minor comments (1)

[Abstract] The abstract refers to 'quantized local models' without naming the specific model families, parameter counts, or quantization formats used; this information should be added to the experimental setup for reproducibility.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive and detailed feedback, which highlights important areas for improving the transparency and rigor of our work. We address each major comment below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Methodology, step 4] LLM-as-a-Judge Evaluation (step 4): The headline finding of an inverse relationship between model scale and consistency rests entirely on scores produced by an LLM judge with chain-of-thought. No details are supplied on the identity or size of the judge model, any ablation of alternative judges, or any correlation of the resulting scores against human expert ratings on the same boolean indicators. Without such grounding, systematic preferences in the judge could mechanically generate the reported sycophancy penalty on larger models.

Authors: We agree that additional details on the LLM judge are required for reproducibility and to mitigate concerns about bias. In the revised manuscript, we will explicitly specify the identity, size, and prompting configuration of the judge model (a distinct quantized model chosen for its balance of capability and consistency). We will also report any ablations on alternative judges conducted during development. The semantic filtering stage and boolean indicator design were intended to reduce subjectivity, and the chain-of-thought prompting was used to increase interpretability of scores. However, we did not perform a human correlation study in the original work. We will add a dedicated limitations subsection discussing potential judge biases and the observed scale-consistency pattern, while noting that results were consistent across all five scenarios. revision: partial
Referee: [Experiments and Results] Experimental results across five scenarios: The manuscript asserts that the results demonstrate subtle trade-offs and the inverse scale-consistency relationship, yet the provided text contains no tables of per-model scores, error bars, statistical significance tests, exclusion criteria, or raw indicator-level data. This absence prevents assessment of whether the 8-9B advantage is reliable or driven by a small number of scenarios.

Authors: We apologize for the insufficient detail in the results presentation. The revised manuscript will include new tables reporting per-model scores across all five legal scenarios, with error bars, statistical significance tests (e.g., paired t-tests or ANOVA for scale comparisons), and explicit exclusion criteria. Raw indicator-level data will be added to an appendix. These additions will allow independent assessment of whether the 8-9B advantage is robust or scenario-specific. revision: yes

standing simulated objections not resolved

Providing a direct correlation between LLM judge scores and human expert ratings on the boolean indicators, as this would require a new human evaluation study not included in the original experiments.

Circularity Check

0 steps flagged

No circularity: framework and results are independently described

full rationale

The paper introduces RPA-Check as a novel four-stage pipeline (dimension definition, augmentation, semantic filtering, LLM-as-Judge) and reports experimental outcomes on five legal scenarios without any equations, fitted parameters renamed as predictions, or load-bearing self-citations. The inverse scale-consistency finding is presented as an empirical observation from applying the pipeline to quantized models, not derived by construction from prior author work or internal definitions. No self-definitional loops, ansatz smuggling, or renaming of known results appear in the abstract or described methodology. The derivation chain is self-contained as a methodological proposal validated by application.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim depends on the reliability of LLM-as-judge scoring and the effectiveness of semantic filtering to produce objective indicators; these are domain assumptions rather than derived results.

axioms (2)

domain assumption LLM-as-a-Judge equipped with chain-of-thought reasoning can produce objective fidelity scores for role-playing behavior
Step 4 of the pipeline relies on this to generate the final evaluation scores.
ad hoc to paper Semantic filtering can eliminate redundancy and ensure agent isolation in the boolean checklist without losing coverage of role-adherence dimensions
Step 3 explicitly invokes this to guarantee indicator quality.

invented entities (1)

RPA-Check four-stage pipeline no independent evidence
purpose: To provide an automated, objective evaluation metric for dynamic LLM role-playing agents
The framework itself is the primary new construct introduced by the paper.

pith-pipeline@v0.9.0 · 5586 in / 1601 out tokens · 86656 ms · 2026-05-10T15:13:11.509421+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 2 internal anchors

[1]

URL:https://arxiv.org/abs/2411.02714,arXiv:2411.02714

Game plot design with an llm-powered assistant: An empirical study with game designers. URL:https://arxiv.org/abs/2411.02714,arXiv:2411.02714. Ashby, T., Webb, B.K., Knapp, G., Searle, J., Fulda, N.,

work page arXiv
[2]

Personalized quest and dialogue generation in role-playing games: A knowledge graph-andlanguagemodel-basedapproach,in:Proceedingsofthe2023CHIConferenceonHumanFactorsinComputingSystems,Association for Computing Machinery, New York, NY, USA. pp. 1–20. doi:10.1145/3544548.3581441. Bai, G., Chai, Z., Ling, C., Wang, S., Lu, J., Zhang, N., Shi, T., Yu, Z., Zhu...

work page doi:10.1145/3544548.3581441
[3]

URL:https://arxiv.org/abs/2401.00625, arXiv:2401.00625

Beyond efficiency: A systematic survey of resource-efficient large language models. URL:https://arxiv.org/abs/2401.00625, arXiv:2401.00625. Beattie,S.,Colbran,S.,2017. Fromphoenixwrighttoatticusfinch:Legalsimulationgamesasanaidtoself-representedlitigants,in:Proceedingsof the 26th International Conference on World Wide Web Companion, International World Wi...

work page doi:10.1145/3041021.3054171 2017
[4]

Using llms to adapt serious games with educators in the loop, in: Schönbohm, A., Bellotti, F., Bucchiarone,A.,deRosa,F.,Ninaus,M.,Wang,A.,Wanick,V.,Dondio,P.(Eds.),GamesandLearningAlliance,SpringerNatureSwitzerland, Cham. pp. 68–77. doi:10.1007/978-3-031-78269-5_7. Chen,J.,Wang,X.,Xu,R.,Yuan,S.,Zhang,Y.,Shi,W.,Xie,J.,Li,S.,Yang,R.,Zhu,T.,Chen,A.,Li,N.,Che...

work page doi:10.1007/978-3-031-78269-5_7
[5]

Chatbot arena: an open platform for evaluating llms by human preference, in: Proceedings of the 41st International Conference on Machine Learning, JMLR.org. pp. 8359–8388. doi:10.5555/3692070.3692401. Cox, S.R., Ooi, W.T.,

work page doi:10.5555/3692070.3692401
[6]

(Eds.), Chatbot Research and Design, Springer Nature Switzerland, Cham

Conversational interactions with npcs in llm-driven gaming: Guidelines from a content analysis of player feedback, in: Følstad, A., Araujo, T., Papadopoulos, S., Law, E.L.C., Luger, E., Goodwin, M., Hobert, S., Brandtzaeg, P.B. (Eds.), Chatbot Research and Design, Springer Nature Switzerland, Cham. pp. 167–184. doi:10.1007/978-3-031-54975-5_10. Fabbri,A.R...

work page doi:10.1007/978-3-031-54975-5_10 2021
[7]

Procedural content generation in games: A survey with insights on emerging llm integration, in: Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, pp. 167–178. doi:10.1609/aiide.v20i1.31877. Gopalakrishnan, K., Hedayatnia, B., Chen, Q., Gottardi, A., Kwatra, S., Venkatesh, A., Gabriel, R., Hakkani-Tur, D.,

work page doi:10.1609/aiide.v20i1.31877
[8]

URL:https://arxiv.org/abs/2308.11995,arXiv:2308.11995

Topical-chat: Towards knowledge-grounded open-domain conversations. URL:https://arxiv.org/abs/2308.11995,arXiv:2308.11995. R. Rosati et al.:Preprint submitted to ElsevierPage 32 of 34 RPA-Check: A Multi-Stage Automated Framework for Evaluating Dynamic LLM-based Role-Playing Agents Gu,J.,Jiang,X.,Shi,Z.,Tan,H.,Zhai,X.,Xu,C.,Li,W.,Shen,Y.,Ma,S.,Liu,H.,Wang,...

work page arXiv
[9]

The Innovation , 101253doi:10.1016/j.xinn.2025.101253

A survey on llm-as-a-judge. The Innovation , 101253doi:10.1016/j.xinn.2025.101253. Gursesli, M.C., Taveekitworachai, P., Abdullah, F., Dewantoro, M.F., Lanata, A., Guazzini, A., Lê, V.K., Villars, A., Thawonmas, R.,

work page doi:10.1016/j.xinn.2025.101253 2025
[10]

(Eds.), Interactive Storytelling, Springer Nature Switzerland, Cham

The chronicles of chatgpt: Generating and evaluating visual novel narratives on climate change through chatgpt, in: Holloway-Attaway, L., Murray, J.T. (Eds.), Interactive Storytelling, Springer Nature Switzerland, Cham. pp. 181–194. doi:10.1007/978-3-031-47658-7_16. Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., Steinhardt, J.,

work page doi:10.1007/978-3-031-47658-7_16
[11]

Measuring Massive Multitask Language Understanding

Measuring massive multitask language understanding. URL:https://arxiv.org/abs/2009.03300,arXiv:2009.03300. Hua, M., Raley, R.,

work page internal anchor Pith review Pith/arXiv arXiv 2009
[12]

Scenecraft: Automating interactive narrative scene generation in digital games with large language models, in: Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, pp. 86–96. doi:10.1609/aiide.v19i1.27504. Lang, J., Guo, Z., Huang, S.,

work page doi:10.1609/aiide.v19i1.27504
[13]

A comprehensive study on quantization techniques for large language models, in: 2024 4th International Conference on Artificial Intelligence, Robotics, and Communication (ICAIRC), pp. 224–231. doi:10.1109/ICAIRC64177.2024.10899941. Lanzi, P.L.,Loiacono, D.,2023. Chatgpt andother largelanguage modelsas evolutionaryengines foronline interactivecollaborative...

work page doi:10.1109/icairc64177.2024.10899941 2024
[14]

Orak: A Foundational Benchmark for Training and Evaluating LLM Agents on Diverse Video Games

Orak: A foundational benchmark for training and evaluating llm agents on diverse video games. URL: https://arxiv.org/abs/2506.03610,arXiv:2506.03610. Pereira, J., Assumpcao, A., Lotufo, R.,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh

Check-eval: A checklist-based approach for evaluating text quality. URL:https://arxiv.org/ abs/2407.14467,arXiv:2407.14467. Phan, L., Mazeika, M., Zou, A., Hendrycks, D.,

work page arXiv
[16]

Textquests: How good are llms at text-based video games?, 2025

Textquests: How good are llms at text-based video games? URL:https://arxiv.org/ abs/2507.23701,arXiv:2507.23701. Salinas, A., Morstatter, F.,

work page arXiv
[17]

The butterfly effect of altering prompts: How small changes and jailbreaks affect large language model performance,in:Ku,L.W.,Martins,A.,Srikumar,V.(Eds.),FindingsoftheAssociationforComputationalLinguistics:ACL2024,Association for Computational Linguistics, Bangkok, Thailand. pp. 4629–4651. doi:10.18653/v1/2024.findings-acl.275. SavannaDevelopments,2024.V...

work page doi:10.18653/v1/2024.findings-acl.275 2024
[18]

Quantifying language models’ sensitivity to spurious features in prompt design or: how i learned to start worrying about prompt formatting, in: 12th International Conference on Learning Representations, ICLR 2024, pp. 1–29. R. Rosati et al.:Preprint submitted to ElsevierPage 33 of 34 RPA-Check: A Multi-Stage Automated Framework for Evaluating Dynamic LLM-...

work page arXiv 2024
[19]

learningversevr

Developing an immersive game-based learning platform with generative artificial intelligence and virtual reality technologies – “learningversevr”. Computers & Education: X Reality 4, 100069. doi:10.1016/j.cexr.2024.100069. Sweetser, P.,

work page doi:10.1016/j.cexr.2024.100069 2024
[20]

Large language models and video games: A preliminary scoping review, in: Proceedings of the 6th ACM Conference on Conversational User Interfaces, Association for Computing Machinery, New York, NY, USA. pp. 1–8. doi:10.1145/3640794.3665582. Tódová,T.,2025. Aquestforinformation:Enhancinggame-basedlearningwithllm-drivennpcs,in:ProceedingsofCESCG2025:The29thC...

work page doi:10.1145/3640794.3665582 2025
[21]

(Eds.), Interactive Storytelling, Springer Nature Switzerland, Cham

Llm-powered npcs, in: Reyes, M.C., Nack, F. (Eds.), Interactive Storytelling, Springer Nature Switzerland, Cham. pp. 171–189. doi:10.1007/978-3-032-12405-0_10. Wang, Y., Chen, J., Xiao, H.,

work page doi:10.1007/978-3-032-12405-0_10
[22]

URL: https://arxiv.org/abs/2601.10122,arXiv:2601.10122

Role-playing agents driven by large language models: Current status, challenges, and future trends. URL: https://arxiv.org/abs/2601.10122,arXiv:2601.10122. Wei,J.,Wang,X.,Schuurmans,D.,Bosma,M.,Ichter,B.,Xia,F.,Chi,E.H.,Le,Q.V.,Zhou,D.,2022. Chain-of-thoughtpromptingelicitsreasoning in large language models, in: Proceedings of the 36th International Confe...

work page doi:10.5555/3600270.3602070 2022
[23]

IEEE Transactions on Games , 1–15doi:10.1109/TG.2025.3564869

A systematic review of generative ai on game character creation: Applications, challenges, and future trends. IEEE Transactions on Games , 1–15doi:10.1109/TG.2025.3564869. Yang, T., Zhu, Y., Quan, X., Liu, C., Wang, Q.,

work page doi:10.1109/tg.2025.3564869 2025
[24]

arXiv preprint arXiv:2502.03821 , year=

Psyplay: Personality-infused role-playing conversational agents. URL:https://arxiv. org/abs/2502.03821,arXiv:2502.03821. Zargham, N., Tonini, L., Alexandrovsky, D., Ruthven, E.G., Friehs, M.A., Dratzidis, L.T., Dänekas, B., Bikas, I., Nacke, L.E., Zebel, S., Malaka, R.,

work page arXiv
[25]

International Journal of Human–Computer Interaction , 1–29doi:10.1080/10447318.2026.2620647

Dialogs with genai npcs: Exploring player interactions with speech agents in a vr game. International Journal of Human–Computer Interaction , 1–29doi:10.1080/10447318.2026.2620647. Zhao, Y., Pan, J., Dong, Y., Dong, T., Wang, G., Ying, F., Shen, Q., Cao, J.,

work page doi:10.1080/10447318.2026.2620647 2026
[26]

Language urban odyssey: A serious game for enhancing second language acquisition through large language models, in: Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, Association for Computing Machinery, New York, NY, USA. pp. 1–7. doi:10.1145/3613905.3651112. Zhu, Q., Zhao, R., Liang, B., Du, J., Gui, L., He, Y.,

work page doi:10.1145/3613905.3651112
[27]

Player*: Enhancing llm-based multi-agent communication and interaction in murder mystery games

Player*: Enhancing llm-based multi-agent communication and interaction in murder mystery games. URL:https://arxiv.org/abs/2404.17662,arXiv:2404.17662. Zhu, X., Chen, Y., Tian, H., Tao, C., Su, W., Yang, C., Huang, G., Li, B., Lu, L., Wang, X., Qiao, Y., Zhang, Z., Dai, J.,

work page arXiv
[28]

Ghost in the minecraft: Generally capable agents for open-world enviroments via large language models with text-based knowledge and memory

Ghost in the minecraft: Generally capable agents for open-world environments via large language models with text-based knowledge and memory. URL: https://arxiv.org/abs/2305.17144,arXiv:2305.17144. Özkaya, S., Berrezueta-Guzman, S., Wagner, S.,

work page arXiv
[29]

IEEE Access 13, 193335–193355

How llms are shaping the future of virtual reality. IEEE Access 13, 193335–193355. doi:10.1109/ACCESS.2025.3631594. R. Rosati et al.:Preprint submitted to ElsevierPage 34 of 34

work page doi:10.1109/access.2025.3631594 2025