######## v1 -> v2
only in v1 []
only in v2 ['Tables/case_critics.tex', 'Tables/exp_diff_judges.tex', 'Tables/human_eval.tex', 'Tables/judge_alignment.tex']
-- Sections/0_Abstract.tex: 1 changes
  - v1:11: Our code, dataset and models are available at \href{https://github.com/Neph0s/CoSER}{https://github.com/Neph0s/CoSER}.
  + v2:11: Our code, dataset and models are available at: \href{https://github.com/Neph0s/CoSER}{https://github.com/Neph0s/CoSER}.
-- Sections/1_Introduction.tex: 1 changes
  - v1:65: \item Results of extensive experiments demonstrate the significant value of our dataset for the training, retrieval and evaluation of RPLAs. Notably, the \method models achieve state-of-the-art performance across four benchmarks for RPLAs.
  + v2:65: \item Results of extensive experiments demonstrate the significant value of our dataset for the training, retrieval and evaluation of RPLAs. Notably, the \method models achieve state-of-the-art performance across four benchmarks and human evaluation for RPLAs.
-- Sections/2_Related_Work.tex: 2 changes
  - v1:5: An RPLA leverages LLMs to create a simulated persona $\agent_\persona$ that emulates a real character $\persona$ based on its persona data $\personadata_\persona$.
  + v2:5: Using LLMs to simulate human personas represents a pioneering research direction.
  + v2:6: Early studies develop prototypes of RPLAs for fictional characters~\citep{shao2023character} and multi-agent systems that simulate human society~\citep{Park2023GenerativeAgents}, while exploring the nature~\citep{shanahan2023role} and potential limitations~\citep{liu-etal-2024-evaluating-large, deshpande-etal-2023-toxicity} of LLM role-play.
  + v2:7: ~\citet{chen2024from} present a comprehensive survey of relevant research.
  + v2:9: Specifically, an RPLA leverages LLMs to create a simulated persona $\agent_\persona$ that simulates a real character $\persona$ based on its persona data $\personadata_\persona$.
  - v1:13: \textit{1)} Many are synthesized via LLMs' responses to general instruction sets~\citep{wang2023rolellm} or character-specific questions~\citep{shao2023character}, such as RoleBench~\citep{wang2023rolellm}.
  + v2:17: \textit{1)} Many are synthesized via LLMs' responses to general instruction sets~\citep{wang2023rolellm, chan2024personahub} or character-specific questions~\citep{shao2023character}, such as RoleBench~\citep{wang2023rolellm}.
== Sections/3_Dataset.tex: identical after whitespace normalisation
-- Sections/4_Methodology.tex: 2 changes
  - v1:76: The initial score for each dimension is calculated as $s = 100-\sum_{f\in\mathcal{F}}v_f$.
  - v1:78: The rubrics are derived from human-annotated issues in extensive human-RPLA conversations from real users and our multi-agent simulations.
  - v1:79: We also consider evaluation dimensions from previous work~\citep{shanahan2023role, shao2023character, chen2024from}.
  + v2:76: The initial score for each dimension is calculated as $s = 100-5*\sum_{f\in\mathcal{F}}v_f$.
  + v2:78: \begin{figure}[h]
  + v2:79: \centering
  + v2:80: \includegraphics[width=0.5\linewidth]{Figures/CoSER-dims.pdf}
  + v2:81: \caption{
  + v2:82: We divide the RPLA evaluation into four quadrants (dimensions).
  + v2:83: The X-axis represents the evaluation perspective: Referenced (comparing with data in the source book) versus Inherent (assessing standalone quality).
  + v2:84: The Y-axis indicates the evaluation scope: Individual Agent versus Overall Simulation.
  + v2:85: }
  + v2:86: \label{fig:dims}
  + v2:87: \end{figure}
  + v2:89: We consider four evaluation dimensions, as shown in Fig. ~\ref{fig:dims}.
  + v2:90: The rubrics are derived from human-annotated issues in extensive human-RPLA conversations from real users and our multi-agent simulations, as well as dimensions from previous work~\citep{shanahan2023role, shao2023character, chen2024from}.
  - v1:92: $s = 100-\sum_{f\in\mathcal{F}}v_f + \lambda|\bar{M}|$,
  + v2:103: $s = 100-5*\sum_{f\in\mathcal{F}}v_f + \lambda|\bar{M}|$,
-- Sections/5_Experiments.tex: 2 changes
  + v2:64: \subsection{Human Evaluation}
  + v2:65: \label{sec:human_eval}
  + v2:67: We conduct human evaluation
  + v2:68: to further validate the performance of \method models and the reliability of LLM critics in GCA evaluation.
  + v2:70: \paragraph{Annotations}
  + v2:71: We randomly select 60 samples from \method Test, and evaluate 7 representative LLMs (shown in Table~\ref{tab:human_eval}) based on their simulation results.
  + v2:72: Three annotators independently score the LLMs on 20 samples each, using a 1 to 10 scoring scale, given the background context and original conversations. During this process, the annotators note that
  + v2:73: \textit{1)} GCA simulation deeply reflects LLMs' role-playing abilities, compared with previous single-turn benchmarks; \textit{2)} Manual evaluation is highly challenging and time-consuming. It requires careful learning of complex background and abundant dialogues, and takes 15 minutes to evaluate 7 models for one sample on average.
  + v2:74: We report models' average scores and win rate against other models.
  + v2:76: \input{Tables/human_eval}
  + v2:78: \paragraph{Results and Analysis}
  + v2:79: According to Table~\ref{tab:human_eval}, we find that:
  + v2:80: \textit{1)} The results generally align with GCA evaluation, validating the reliability of our LLM-based evaluation protocol;
  + v2:81: \textit{2)} \method 70B maintain superior performance, obtaining the highest score (6.783) and win rate (86.9\%);
  + v2:82: \textit{3)} Human evaluators show less preference for GPT models compared to GPT-4o as LLM judges, which likely stems from self-preference bias in LLM judges~\citep{wataoka2024self}.
  + v2:85: \paragraph{Alignment}
  + v2:86: We further study the alignment between LLM and human judges.
  + v2:87: We evaluate three judge models, GPT-4o, DeepSeek-V3 and DeepSeek-R1~\citep{guo2025deepseek}.
  + v2:88: We consider standard GCA evaluation and ablation variants that remove: \textit{(i)} original dialogues as reference, \textit{(ii)} scoring rubrics, \textit{(iii)} length correction, and \textit{(iv)} dimension separation.
  + v2:89: We also compare with BLEU and ROUGE-L.
  + v2:90: The alignment is measured via model-wise preferences (\S\ref{sec:appendix_human_eval}).
  + v2:92: The results in Table~\ref{tab:judge_alignment} reveal that:
  + v2:93: \textit{1)} All components of GCA evaluation contribute to improved alignment;
  + v2:94: \textit{2)} Reasoning models excel as judges, with DeepSeek-R1 achieving the highest alignment;
  + v2:95: \textit{3)} BLEU and ROUGE-L remain effective indicators for role-playing evaluation.
  + v2:97: \input{Tables/judge_alignment}
  - v1:105: \subsection{Case studies}
  + v2:140: \subsection{Case Studies}
== Sections/7_Conclusion.tex: identical after whitespace normalisation
-- Sections/Appendix.tex: 7 changes
  - v1:12: On average, each conversation consists of approximately 13.2 utterances, with the entire dataset comprising a total of 392,900 utterances.
  + v2:12: On average, each conversation consists of approximately 13.2 utterances, with the entire dataset comprising a total of 392,298 utterances.
  - v1:19: 771 & 30,069 & 29,798 & 17,966 & 392,900 \\ \hline
  + v2:20: 771 & 30,069 & 29,798 & 17,966 & 392,298 \\ \hline
  + v2:88: \subsection{Safety}
  + v2:89: We conduct safety checks on the dataset and remove potentially problematic content. Specifically, we truncate 110 sensitive conversations and remove a total of 602 messages.
  + v2:91: \subsection{Comparison with PersonaHub}
  + v2:93: Recently, several studies have explored LLM role-playing using synthesized personas rather, such as PersonaHub~\citep{chan2024personahub}. We provide a systematic comparison between \method and PersonaHub below.
  + v2:95: \paragraph{Goal and Focus}
  + v2:96: The fundamental objectives of the two datasets differ significantly:
  + v2:98: \begin{table}[h]
  + v2:99: \centering
  + v2:100: \begin{tabular}{lp{5cm}p{5cm}}
  + v2:101: \hline
  + v2:102: \textbf{Aspect} & \textbf{\method} & \textbf{PersonaHub} \\
  + v2:103: \hline
  + v2:104: Target & Simulate established personas with high fidelity & Synthesize instruction data for knowledge distillation \\
  + v2:105: \hline
  + v2:106: Persona Focus & Depth and richness of character data & Breadth of persona types \\
  + v2:107: \hline
  + v2:108: Key Capabilities & Anthropomorphism, character fidelity, multi-character interaction & General instruction following \\
  + v2:109: \hline
  + v2:110: \end{tabular}
  + v2:111: \caption{Comparison between \method and PersonaHub regarding goal and focus.}
  + v2:112: \end{table}
  + v2:114: \paragraph{Dataset Quality and Design}
  + v2:115: The data sources and design reflect distinct research priorities:
  + v2:117: \begin{table}[h]
  + v2:118: \centering
  + v2:119: \begin{tabular}{lp{5cm}p{5cm}}
  + v2:120: \hline
  + v2:121: \textbf{Aspect} & \textbf{\method} & \textbf{PersonaHub} \\
  + v2:122: \hline
  + v2:123: Data Source & Authentic book dialogues & GPT-4o synthesized conversations \\
  + v2:124: \hline
  + v2:125: Persona Origin & Book characters & LLM-synthesized personas \\
  + v2:126: \hline
  + v2:127: Persona Data & Comprehensive profiles, dialogues, experiences & Concise profiles with limited context \\
  + v2:128: \hline
  + v2:129: Design Priority & Quality, authenticity, and character depth & Quantity and diversity\\
  + v2:130: \hline
  + v2:131: \end{tabular}
  + v2:132: \caption{Comparison between \method and PersonaHub regarding dataset quality and design.}
  + v2:133: \end{table}
  + v2:136: This comparison highlights that while PersonaHub excels in generating diverse synthetic personas for broad instruction-following training, \method focuses on preserving the authenticity and depth of established characters for high-fidelity role-playing applications.
  - v1:173: \section{Examples and Case Study}
  - v1:175: We present several examples of our extracted conversations, as well as corresponding simulations in this given circumstance by LLMs.
  - v1:177: Tables~\ref{tab:case_sansa} to~\ref{tab:case_sansa_2} illustrate a classic conversation extracted from \textit{A Storm of Swords (A Song of Ice and Fire, \#3)} and the corresponding simulation by \method 70B. In \method 70B’s simulation, when confronted with \textit{Sansa Stark}, \textit{Lysa Arryn} utters her iconic line, \textit{“Grown enough to be wed, wed enough to be bedded”}, reflecting her personality and worldview.
  - v1:178: This indicates that \method models excellently recall and apply character-related knowledge from their pretrained data.
  - v1:180: Tables \ref{tab:case_cerci} to \ref{tab:case_cerci4} present another scene from \textit{A Dance with Dragons (A Song of Ice and Fire, \#5)}, specifically the \textit{walk of atonement}, in which \textit{Cersei Lannister} is forced to walk naked through the streets, facing both physical and mental humiliation while striving to preserve her dignity.
  - v1:181: We present the original dialogue alongside simulations by \method 70B, GPT-4o, and Claude-3.5-Sonnet.
  - v1:182: Notably, \method 70B faithfully captures the suppressed anger of \textit{Cersei Lannister} as depicted in the original conversation, whereas the other models, including GPT-4o and Claude-3.5-Sonnet, resort to a stereotypical portrayal of her arrogance and pride.
  - v1:185: \input{Tables/case_sansa}
  - v1:187: \input{Tables/case_sansa_2}
  - v1:189: \input{Tables/case_cerci}
  - v1:191: \input{Tables/case_cerci_2}
  - v1:193: \input{Tables/case_cerci_3}
  - v1:195: \input{Tables/case_cerci_4}
  + v2:224: \subsection{Human Evaluation}
  + v2:225: \label{sec:appendix_human_eval}
  + v2:227: The alignment metrics are measured via the frequency of agreement between LLM and human judges when comparing model pairs, excluding cases where judges assign similar scores to both models (\ie, score differences $\leq$ 1 for human judges or $\leq$ 5\% for automatic metrics).
  + v2:233: \paragraph{Evaluation with Different LLMs as Judges}
  + v2:234: We conduct additional experiments using alternative judge models for GCA evaluation.
  + v2:235: We evaluate 7 representative models as shown in Table~\ref{tab:different_judges},
  + v2:236: using three different judge models: GPT-4o, DeepSeek-R1, and DeepSeek-V3. This comparison aims to examine whether our conclusions remain consistent across different LLM judges.
  + v2:238: \input{Tables/exp_diff_judges}
  + v2:240: The results in Table~\ref{tab:different_judges} reveal several important findings:
  + v2:241: \textit{1)} Despite these variations in absolute scores, the relative ranking patterns remain largely consistent, with \method models maintaining competitive or superior performance across all settings.
  + v2:242: \textit{2)} Different judges exhibit distinct preferences, confirming potential bias in LLM-based evaluation. While GPT-4o shows preference for GPT models, DeepSeek-R1 and DeepSeek-V3 favor \method-70B and Claude-3.5-Sonnet over GPT-4o.
  + v2:296: \paragraph{Comparing GCA Simulation with AI-driven Storytelling Methods}
  + v2:298: To provide a more comprehensive evaluation against prior work in digital actors and AI-driven storytelling, we compare our GCA simulation approach with HollmWood~\citep{chen-etal-2024-hollmwood}, a representative method for AI-generated storytelling.
  + v2:299: We conduct this comparison using 30 samples from \method Test, employing GPT-4o as both the actor and writer in both frameworks to ensure fair comparison.
  + v2:300: The generated stories and character interactions are evaluated using GCA evaluation.
  + v2:301: Results demonstrate that GCA simulation achieves an average score of 59.2\%, significantly outperforming HollmWood's 50.2\%.
  + v2:302: These findings highlight that our given-circumstance acting approach produces more authentic and human-like character interactions compared to previous AI storytelling methods.
  + v2:303: The superior performance of GCA simulation can be attributed to its focus on character-driven dialogue generation within well-defined circumstances, rather than plot-driven narrative construction, leading to more faithful character portrayals and realistic conversational dynamics.
  + v2:306: \section{Authentic v.s. LLM-Generated Role-playing Data}
  + v2:308: To demonstrate the superiority of authentic role-playing data over synthesized alternatives, we conduct two complementary experiments comparing authentic conversations from \method dataset with LLM-generated dialogues.
  + v2:310: \paragraph{GCA Evaluation of Dialogue Quality}
  + v2:311: We apply GCA evaluation to score authentic (groundtruth) and LLM-generated dialogues on 100 samples from \method Test without providing groundtruth to LLM judges as reference.
  + v2:312: We use DeepSeek-R1 as the judge model due to its superior alignment with human evaluators (as shown in \S\ref{sec:human_eval}).
  + v2:313: GPT-4o is used to generate synthetic dialogues via GCA simulation.
  + v2:314: Results show that authentic dialogues from \method achieve a score of 85.1\%, significantly outperforming GPT-4o generated dialogues at 76.2\%.
  + v2:316: \paragraph{Training Effectiveness Comparison}
  + v2:317: We further examine the training effectiveness of authentic versus synthetic dialogues.
  + v2:318: We fine-tune LLaMA-3.1-8B using authentic or LLM-generated dialogues for 200 samples from \method Test for 4 epochs, using the same settings as \method-8B.
  + v2:319: We evaluate the two trained models via GCA evaluation on 100 new samples.
  + v2:320: The model trained on authentic dialogues from \method achieves 55.3\% score, while the model trained on GPT-4o generated dialogues achieves 54.7\%.
  + v2:322: These results demonstrate that authentic dialogues consistently outperform LLM-generated alternatives in both intrinsic quality and training effectiveness.
  + v2:323: These findings highlight the importance of authentic, high-quality role-playing data for developing effective RPLAs.
  + v2:325: \section{Examples and Case Study}
  + v2:327: We present several examples of our extracted conversations, as well as corresponding simulations in this given circumstance by LLMs.
  + v2:329: Tables~\ref{tab:case_sansa} to~\ref{tab:case_sansa_2} illustrate a classic conversation extracted from \textit{A Storm of Swords (A Song of Ice and Fire, \#3)} and the corresponding simulation by \method 70B. In \method 70B’s simulation, when confronted with \textit{Sansa Stark}, \textit{Lysa Arryn} utters her iconic line, \textit{“Grown enough to be wed, wed enough to be bedded”}, reflecting her personality and worldview.
  + v2:330: This indicates that \method models excellently recall and apply character-related knowledge from their pretrained data.
  + v2:332: Tables \ref{tab:case_cerci} to \ref{tab:case_cerci4} present another scene from \textit{A Dance with Dragons (A Song of Ice and Fire, \#5)}, specifically the \textit{walk of atonement}, in which \textit{Cersei Lannister} is forced to walk naked through the streets, facing both physical and mental humiliation while striving to preserve her dignity.
  + v2:333: We present the original dialogue alongside simulations by \method 70B, GPT-4o, and Claude-3.5-Sonnet.
  + v2:334: Notably, \method 70B faithfully captures the suppressed anger of \textit{Cersei Lannister} as depicted in the original conversation, whereas the other models, including GPT-4o and Claude-3.5-Sonnet, resort to a stereotypical portrayal of her arrogance and pride.
  + v2:336: Table~\ref{tab:critique_example} presents an output example of LLM critics, demonstrating the identified performance flaws in GCA simulation.
  + v2:338: \input{Tables/case_sansa}
  + v2:340: \input{Tables/case_sansa_2}
  + v2:342: \input{Tables/case_cerci}
  + v2:344: \input{Tables/case_cerci_2}
  + v2:346: \input{Tables/case_cerci_3}
  + v2:348: \input{Tables/case_cerci_4}
  + v2:350: \input{Tables/case_critics}
  - v1:286: Despite our penalty-based scoring mechanism and detailed rubrics, problems such as length bias persist~\citep{}. Moreover, LLM Judges may lack the necessary knowledge to accurately evaluate character fidelity.
  + v2:384: Despite our penalty-based scoring mechanism and detailed rubrics, problems such as length bias persist. Moreover, LLM Judges may lack the necessary knowledge to accurately evaluate character fidelity.
== Tables/case_cerci.tex: identical after whitespace normalisation
== Tables/case_cerci_2.tex: identical after whitespace normalisation
== Tables/case_cerci_3.tex: identical after whitespace normalisation
== Tables/case_cerci_4.tex: identical after whitespace normalisation
== Tables/case_sansa.tex: identical after whitespace normalisation
== Tables/case_sansa_2.tex: identical after whitespace normalisation
== Tables/dataset_features.tex: identical after whitespace normalisation
== Tables/exp_idood.tex: identical after whitespace normalisation
== Tables/exp_main_cot_full.tex: identical after whitespace normalisation
== Tables/exp_otherbmks.tex: identical after whitespace normalisation
== Tables/exp_rag.tex: identical after whitespace normalisation
== Tables/exp_wr.tex: identical after whitespace normalisation
== Tables/main_exp.tex: identical after whitespace normalisation
== Tables/main_exp_cf3.tex: identical after whitespace normalisation
== Tables/main_exp_cot.tex: identical after whitespace normalisation
== Tables/prompts_agent.tex: identical after whitespace normalisation
== Tables/prompts_agent_2.tex: identical after whitespace normalisation
== Tables/prompts_data.tex: identical after whitespace normalisation
== Tables/prompts_data_2.tex: identical after whitespace normalisation
== Tables/prompts_data_3.tex: identical after whitespace normalisation
== Tables/prompts_eval.tex: identical after whitespace normalisation
== Tables/prompts_eval_2.tex: identical after whitespace normalisation
== Tables/top_books.tex: identical after whitespace normalisation
-- main.bbl: 10 changes
  - v1:1: \begin{thebibliography}{37}
  + v2:1: \begin{thebibliography}{43}
  + v2:8: \bibitem[Chan et~al.(2024)Chan, Wang, Yu, Mi, and Yu]{chan2024personahub}
  + v2:9: Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, and Dong Yu.
  + v2:10: \newblock Scaling synthetic data creation with 1,000,000,000 personas.
  + v2:11: \newblock \emph{ArXiv preprint}, abs/2406.20094, 2024.
  + v2:12: \newblock URL \url{https://arxiv.org/abs/2406.20094}.
  + v2:34: \bibitem[Chen et~al.(2024{\natexlab{c}})Chen, Zhu, Yang, Shi, Xi, Zhang, Wang, Pu, Feng, Yang, and Zhang]{chen-etal-2024-hollmwood}
  + v2:35: Jing Chen, Xinyu Zhu, Cheng Yang, Chufan Shi, Yadong Xi, Yuxiang Zhang, Junjie Wang, Jiashu Pu, Tian Feng, Yujiu Yang, and Rongsheng Zhang.
  + v2:36: \newblock {H}o{LLM}wood: Unleashing the creativity of large language models in screenwriting via role playing.
  + v2:37: \newblock In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, \emph{Findings of the Association for Computational Linguistics: EMNLP 2024}, pages 8075--8121, Miami, Florida, USA, November 2024{\natexlab{c}}. Association for Computational Linguistics.
  + v2:38: \newblock \doi{10.18653/v1/2024.findings-emnlp.474}.
  + v2:39: \newblock URL \url{https://aclanthology.org/2024.findings-emnlp.474/}.
  + v2:53: \bibitem[Deshpande et~al.(2023)Deshpande, Murahari, Rajpurohit, Kalyan, and Narasimhan]{deshpande-etal-2023-toxicity}
  + v2:54: Ameet Deshpande, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, and Karthik Narasimhan.
  + v2:55: \newblock Toxicity in chatgpt: Analyzing persona-assigned language models.
  + v2:56: \newblock In Houda Bouamor, Juan Pino, and Kalika Bali, editors, \emph{Findings of the Association for Computational Linguistics: EMNLP 2023}, pages 1236--1270, Singapore, 2023. Association for Computational Linguistics.
  + v2:57: \newblock \doi{10.18653/v1/2023.findings-emnlp.88}.
  + v2:58: \newblock URL \url{https://aclanthology.org/2023.findings-emnlp.88}.
  + v2:78: \bibitem[Guo et~al.(2025)Guo, Yang, Zhang, Song, Zhang, Xu, Zhu, Ma, Wang, Bi, et~al.]{guo2025deepseek}
  + v2:79: Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et~al.
  + v2:80: \newblock Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.
  + v2:81: \newblock \emph{arXiv preprint arXiv:2501.12948}, 2025.
  - v1:101: \bibitem[Liu et~al.(2024)Liu, Feng, Xue, Wang, Wu, Lu, Zhao, Deng, Zhang, Ruan, et~al.]{liu2024deepseek}
  + v2:126: \bibitem[Liu et~al.(2024{\natexlab{a}})Liu, Feng, Xue, Wang, Wu, Lu, Zhao, Deng, Zhang, Ruan, et~al.]{liu2024deepseek}
  - v1:104: \newblock \emph{ArXiv preprint}, abs/2412.19437, 2024.
  + v2:129: \newblock \emph{ArXiv preprint}, abs/2412.19437, 2024{\natexlab{a}}.
  + v2:132: \bibitem[Liu et~al.(2024{\natexlab{b}})Liu, Diab, and Fried]{liu-etal-2024-evaluating-large}
  + v2:133: Andy Liu, Mona Diab, and Daniel Fried.
  + v2:134: \newblock Evaluating large language model biases in persona-steered generation.
  + v2:135: \newblock In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, \emph{Findings of the Association for Computational Linguistics: ACL 2024}, pages 9832--9850, Bangkok, Thailand, August 2024{\natexlab{b}}. Association for Computational Linguistics.
  + v2:136: \newblock \doi{10.18653/v1/2024.findings-acl.586}.
  + v2:137: \newblock URL \url{https://aclanthology.org/2024.findings-acl.586/}.
  + v2:216: \bibitem[Wataoka et~al.(2024)Wataoka, Takahashi, and Ri]{wataoka2024self}
  + v2:217: Koki Wataoka, Tsubasa Takahashi, and Ryokan Ri.
  + v2:218: \newblock Self-preference bias in llm-as-a-judge.
  + v2:219: \newblock \emph{arXiv preprint arXiv:2410.21819}, 2024.
  - v1:216: \newblock URL \url{http://papers.nips.cc/paper\_files/paper/2023/hash/91f18a1287b398d378ef22505bf41832-Abstract-Datasets\_and\_Benchmarks.html}.
  + v2:253: \newblock URL \url{https://openreview.net/forum?id=uccHPGDlao}.
-- main.tex: 4 changes
  + v2:125: \newcommand{\huggingface}{\raisebox{-1.5pt}{\includegraphics[height=1.05em]{Figures/hf-logo.pdf}}\xspace}
  + v2:126: \newcommand{\github}{\raisebox{-1.5pt}{\includegraphics[height=1.05em]{Figures/github-logo.pdf}}\xspace}
  + v2:201: \author[2]{Shuchang Zhou}
  - v1:195: \author[2]{Shuchang Zhou}
  - v1:203: \correspondingauthor{xtwang21@m.fudan.edu.cn. Work done during Xintao Wang's internship at StepFun.}
  + v2:209: \correspondingauthor{Corresponding author(s): xtwang21@m.fudan.edu.cn. Work done during Xintao Wang's internship at StepFun.}
######## v2 -> v3
only in v2 []
only in v3 []
== Sections/0_Abstract.tex: identical after whitespace normalisation
== Sections/1_Introduction.tex: identical after whitespace normalisation
== Sections/2_Related_Work.tex: identical after whitespace normalisation
== Sections/3_Dataset.tex: identical after whitespace normalisation
== Sections/4_Methodology.tex: identical after whitespace normalisation
== Sections/5_Experiments.tex: identical after whitespace normalisation
== Sections/7_Conclusion.tex: identical after whitespace normalisation
== Sections/Appendix.tex: identical after whitespace normalisation
== Tables/case_cerci.tex: identical after whitespace normalisation
== Tables/case_cerci_2.tex: identical after whitespace normalisation
== Tables/case_cerci_3.tex: identical after whitespace normalisation
== Tables/case_cerci_4.tex: identical after whitespace normalisation
== Tables/case_critics.tex: identical after whitespace normalisation
== Tables/case_sansa.tex: identical after whitespace normalisation
== Tables/case_sansa_2.tex: identical after whitespace normalisation
== Tables/dataset_features.tex: identical after whitespace normalisation
== Tables/exp_diff_judges.tex: identical after whitespace normalisation
== Tables/exp_idood.tex: identical after whitespace normalisation
== Tables/exp_main_cot_full.tex: identical after whitespace normalisation
== Tables/exp_otherbmks.tex: identical after whitespace normalisation
== Tables/exp_rag.tex: identical after whitespace normalisation
== Tables/exp_wr.tex: identical after whitespace normalisation
== Tables/human_eval.tex: identical after whitespace normalisation
== Tables/judge_alignment.tex: identical after whitespace normalisation
== Tables/main_exp.tex: identical after whitespace normalisation
== Tables/main_exp_cf3.tex: identical after whitespace normalisation
== Tables/main_exp_cot.tex: identical after whitespace normalisation
== Tables/prompts_agent.tex: identical after whitespace normalisation
== Tables/prompts_agent_2.tex: identical after whitespace normalisation
== Tables/prompts_data.tex: identical after whitespace normalisation
== Tables/prompts_data_2.tex: identical after whitespace normalisation
== Tables/prompts_data_3.tex: identical after whitespace normalisation
== Tables/prompts_eval.tex: identical after whitespace normalisation
== Tables/prompts_eval_2.tex: identical after whitespace normalisation
== Tables/top_books.tex: identical after whitespace normalisation
== main.bbl: identical after whitespace normalisation
-- main.tex: 28 changes
  + v3:2: % \documentclass{article} % For LaTeX2e
  + v3:4: % Optional math commands from https://github.com/goodfeli/dlbook_notation.
  + v3:5: % \input{math_commands.tex}
  + v3:7: % Recommended, but optional, packages for figures and better typesetting:
  + v3:9: %\usepackage{algorithmic}
  + v3:10: % hyperref makes hyperlinks in the resulting PDF.
  + v3:11: % If your build breaks (sometimes temporarily if a hyperlink spans a page)
  + v3:12: % please comment out the following usepackage line and replace
  + v3:13: % \usepackage{icml2023} with \usepackage[nohyperref]{icml2023} above.
  + v3:21: % \linespread{.98}
  + v3:28: % if you use cleveref..
  + v3:45: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:46: % THEOREMS
  + v3:47: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:48: % \theoremstyle{plain}
  + v3:49: % \theoremstyle{definition}
  + v3:50: % \newtheorem{theorem}{Theorem}[section]
  + v3:51: % \newtheorem{proposition}[theorem]{Proposition}
  + v3:52: % \newtheorem{lemma}[theorem]{Lemma}
  + v3:53: % \newtheorem{corollary}[theorem]{Corollary}
  + v3:54: % \newtheorem{definition}[theorem]{Definition}
  + v3:55: % \newtheorem{assumption}[theorem]{Assumption}
  + v3:56: % \theoremstyle{remark}
  + v3:57: % \newtheorem{remark}[theorem]{Remark}
  + v3:59: % Todonotes is useful during development; simply uncomment the next line
  + v3:60: % and comment out the line below the next line to turn off comments
  + v3:61: %\usepackage[disable,textsize=tiny]{todonotes}
  + v3:62: %\usepackage[textsize=tiny]{todonotes}
  - v2:45: \title{\method: Coordinating LLM-Based Persona Simulation of Established Roles}
  + v3:68: % ##########################################
  + v3:70: % ##########################################
  + v3:72: % ##########################################
  + v3:74: % ##########################################
  + v3:76: % ##########################################
  + v3:77: % \documentclass{article} % For LaTeX2e
  + v3:78: % \documentclass[11pt]{berkeley}
  + v3:80: % \usepackage[preprint]{neurips_2023}
  + v3:81: %\title{\method: Coordinating LLM-Based Persona Simulation of Established Roles}
  + v3:83: \title{\method: A Comprehensive Literary Dataset and Framework for Training and Evaluating LLM Role-Playing and Persona Simulation}
  + v3:86: %\title{\method: A Comprehensive Suite for Character Role-Playing of LLMs}
  + v3:91: % \bibliographystyle{plainnat}
  - v2:61: \usepackage{booktabs} %
  + v3:102: \usepackage{booktabs} % for professional tables
  + v3:103: %\usepackage{floatrow}
  + v3:104: %\usepackage{float}
  + v3:116: % \usepackage{minted}
  + v3:117: %\usepackage{float}
  + v3:120: % \setminted[python]{frame=lines, breaklines, framesep=2mm, fontsize=\footnotesize, numbersep=5pt}
  + v3:121: %\usepackage{paralist}
  + v3:123: % \usepackage[authoryear, sort&compress, round]{natbib}
  + v3:143: %\usepackage{paralist}
  + v3:184: % Python style for highlighting
  - v2:138: morekeywords={self, clip, exp, mse_loss, uniform_sample, concatenate, logsumexp}, %
  + v3:188: morekeywords={self, clip, exp, mse_loss, uniform_sample, concatenate, logsumexp}, % Add keywords here
  - v2:140: emph={MyClass,__init__}, %
  - v2:141: emphstyle=\color{deepred}, %
  + v3:190: emph={MyClass,__init__}, % Custom highlighting
  + v3:191: emphstyle=\color{deepred}, % Custom highlighting style
  - v2:143: frame=single, %
  + v3:193: frame=single, % Any extra options here
  + v3:197: % Python environment
  + v3:205: % Python for external files
  + v3:210: % Python for inline
  + v3:215: % % \input{defs.tex}
  + v3:216: % % \input{}
  + v3:228: % \usepackage[textsize=tiny]{todonotes}
  + v3:229: % \setlength{\parskip}{3pt}
  + v3:231: % \usepackage[skins,theorems]{tcolorbox}
  + v3:232: % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:233: % % from Tengyang
  + v3:235: % tweaking cleveref
  + v3:242: % Fix hyperref in section titles.
  + v3:252: % \usepackage[shortlabels]{enumitem}
  + v3:268: %\affil[*]{Equal contributions}
  + v3:288: %\input{Tables/urls}
  + v3:294: %\input{Sections/6_Discussions}
  + v3:299: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:300: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:301: % APPENDIX
  + v3:302: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:303: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:307: % \input{sections/app_frameworksummary}
  + v3:309: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
  + v3:310: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%