Test-Time-Matching: Decouple Personality, Memory, and Linguistic Style in LLM-based Role-Playing Language Agent

Hao Sun; Jie Guo; Xiaoyu Zhan; Xinyu Fu; Yanwen Guo; Yuanqi Li

REVIEW 4 major objections 4 minor 1 cited by

Test-Time-Matching: Decouple Personality, Memory, and Linguistic Style in LLM-based Role-Playing Language Agent

T0 review · 4 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 Test-Time-Matching claims that role-playing fidelity improves when a character's personality, memory, and linguistic style are decoupled and generated in three separate stages.

desk verdict The decoupling claim is the paper's load-bearing wall and it is not tested; the pipeline itself is a solid training-free contribution that deserves review, but the causal story needs ablations before it can be believed. read the letter →

arxiv 2507.16799 v2 pith:OQKPG2J4 submitted 2025-07-22 cs.CL

Xiaoyu Zhan , Xinyu Fu , Hao Sun , Yuanqi Li , Jie Guo , Yanwen Guo This is my paper

classification cs.CL

keywords role-playinglanguageagentstest-timescalingcontextengineeringlinguisticstyletransferpersonalitydecouplingmemoryretrievalconsistencyLLMevaluation

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

The paper proposes Test-Time-Matching (TTM), a training-free framework for role-playing language agents that deliberately separates three components of a character—personality, memory, and linguistic style—and generates dialogue in three stages: a styleless response driven by personality and background, a memory-checked revision using retrieval-augmented generation, and a final style transfer via progressive matching. The claim is that this decoupling, rather than blending all features in one prompt, is what produces higher-fidelity role-playing. Human raters and an LLM judge scored TTM above prompt-only and fine-tuned baselines on persona consistency, knowledge accuracy, and conversation quality. If right, it means style transfer should be a separate final step after content and memory are settled, and characters can be recombined modularly without retraining.

What carries the argument

The load-bearing mechanism is the three-stage pipeline with explicit decoupling: first, a personality and background prompt generates a styleless response; second, that response is rewritten into retrieval keywords for a graph-based retrieval-augmented database, producing a memory-checked response; third, linguistic style, represented as style preferences plus common word categories and retrieved similar utterances, is applied through progressive matching, rewriting sentence by sentence with hybrid retrieval from the character's historical dialogue. The named procedure, Test-Time-Matching, is what carries the argument: by keeping style out of the early stages, the paper claims cognitive content and factual grounding are settled before verbal mannerisms are added.

What would settle it

If an ablation that generates the final response in one combined prompt, or that applies style transfer before memory-checking, achieved equal or better ratings on persona consistency and knowledge accuracy across the same six characters, then the decoupling claim would fail; a more direct test would be to have expert readers identify the target character from the first-stage 'styleless' responses alone—if they can do so reliably, the styleless stage is not style-free.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that an LLM can play a character more faithfully when the character is decomposed into cognitive tendency (personality plus memory) and linguistic style, generated separately, and then recombined. TTM first builds a structured profile—personality traits, background facts, a graph-based memory database, and style features such as preferred word classes and tone—automatically from raw text. At generation time it produces a styleless response, retrieves knowledge to correct and enrich it, and only then applies the character's style using retrieved similar utterances and sentence-by-sentence progressive matching. In the paper's evaluations TTM on Qwen-3-32B scores highest among all methods on persona consistency, accuracy of knowledge, and conversation quality in both LLM-as-judge and general-participant ratings.

Load-bearing premise

The paper's load-bearing premise is that personality and memory can be cleanly separated from linguistic style, so a genuinely styleless response can be generated first and style added later; the paper itself notes that removing style may change the original meaning of a sentence.

Editorial extensions

If this is right

A training-free system can surpass fine-tuned role-players by spending additional compute and context engineering at inference time.
Personality, memory, and style can be swapped independently, allowing controlled customization and recombination across characters.
Applying style as a separate final step improves style consistency: TTM sentences were identified as the target character's with probability 15.48%, versus 24.80% for the base model.
The three-stage pipeline is modular, so memory-checking or style generation can be removed or reordered without retraining.
Test-time methods inherit base model knowledge, so overall performance depends on model capability and pretraining familiarity with the character.

Reading between the lines

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

If the decoupling is robust, role-playing becomes modular: the same personality and memory profile could be paired with any style profile, letting users audition or swap mannerisms without touching the character's knowledge or disposition; the paper demonstrates combinations but does not stress-test this trade-off.
The main confound in the evaluation is the base model's pretraining familiarity with famous characters, so a natural extension is to test TTM on obscure or synthetic characters to isolate what the pipeline adds beyond memorized persona.
Because TTM is training-free and spends extra test-time compute, it suggests a general recipe—generate content, ground memory, then restyle—that could transfer to other generation tasks where tone should be decoupled from content, such as summarization or translation.

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.

Desk Editor's Note

read the letter

What you should know: TTM is a training-free role-playing framework that splits generation into three stages — styleless content, memory-checking via RAG, then style rewrite. That specific combination is new relative to ChatHaruhi, CoSER, StyleChat, and CAT-LLM, and the paper ships code plus a fairly detailed pipeline description. The strong version of the claim, that decoupling is what causes the fidelity gain, is not supported by the experiments.

What the paper does well: the three-stage architecture is coherent and modular; extracting personality, background, and style into explicit plain-text profiles is genuinely useful for controllability; progressive matching and hybrid retrieval are reasonable design choices. The authors are also transparent — the Limitations and Appendix sections openly concede that the base model's pretraining knowledge can drive quality, that the system prompt may hurt some models, that retrieval can cause information overload and knowledge misuse, and that the style-removal step can change meaning. That honesty is to their credit.

The soft spots are real and central. The decoupling claim is never isolated. There are no ablations that remove the style stage, remove memory, or compare against a single-prompt baseline on the same base model. The 'styleless response' is asserted but never shown to be style-free; retrieved chunks can carry style, and the optional style-removal step is used 'with caution' because it may alter meaning. So the measured gain over Qwen-3-32B could come from RAG, longer responses, or extra LLM calls rather than from decoupling. The human evaluation is also thin: 17 raters, one dialogue per character per method, no significance tests. The LLM-as-judge result aligns with the human result on TTM's favor, which is something, but the judge is the same kind of model that does the extraction and rewriting, so there is a mild self-referential flavor. The style identification study is a useful complement, though it uses a weaker base model and the difference is not dramatic.

Who is this for: researchers doing test-time adaptation for role-playing or modular style control. They will get value from the pipeline description, but should treat the decoupling claim as a hypothesis, not a demonstrated result. The paper deserves a serious referee because the idea is novel and the missing experiments are concrete and addressable.

Recommendation: send to peer review, but require ablations that isolate each stage, a single-prompt baseline on the same base model, and at least error bars or significance testing. With those, this could become a solid contribution.

Referee Report

4 major / 4 minor

Summary. The paper proposes Test-Time-Matching (TTM), a training-free role-playing framework that decouples a character's personality, memory, and linguistic style through a three-stage generation pipeline: first, a styleless response is generated from personality and background; second, retrieval-augmented generation (RAG) provides memory-checked details; third, progressive matching rewrites the response to match the target character's extracted linguistic style. The method is evaluated against ChatHaruhi, CoSER, GPT-4o, Gemini-2.5-pro, and Qwen-3 baselines using both LLM-as-Judge and human assessments over three Chinese and three English characters. The reported results show TTM (built on Qwen-3-32B) receiving the highest scores in most dimensions, and the paper claims that explicitly decoupling personality, memory, and linguistic style is what produces this fidelity.

Significance. The paper's core idea — that linguistic style can be cleanly separated from cognitive traits such as personality and memory, enabling modular and controllable role-playing — is interesting and practically relevant. The proposed framework is fully automatic, requires no fine-tuning, and the authors release code, which supports reproducibility. The three-stage pipeline is a novel organizational principle for test-time role-playing, and the inclusion of both general-participant and linguistics-expert human evaluation is a strength relative to LLM-only evaluations. However, the central causal claim that decoupling itself drives the observed gains is not currently supported: the evaluation lacks ablations isolating the stages, the style-neutrality of the first-stage response is never measured, and the human study has a small rater pool with one dialogue per condition and no significance testing. The paper's appendices explicitly concede that base-model pretraining knowledge may drive much of the styleless-phase quality and that style removal is risky, further weakening the decoupling attribution.

major comments (4)

[Experiments, Table 1] The central claim that the three-stage decoupling produces the performance gains is not supported because TTM is only compared as a complete pipeline against baselines. No ablation removes the style-rewrite stage, removes the memory stage, or compares against a single-prompt combined-feature baseline on the same base model. The improvements over Qwen-3-32B could come from RAG, longer responses, or additional LLM calls rather than from decoupling. Please add stage-wise ablations and report per-condition scores on the same base model.
[Method, 'Styleless Response and Memory-checked Response Generation'] The premise that the first-stage response is style-free is never measured. The paper concedes that RAG-retrieved content 'may still contain stylistically marked expressions,' that the optional style-removal step 'may change the original meaning of the sentence,' and that it is used 'with caution.' The appendix section 'The Influences of Base Model' further states that pretraining knowledge of the target character 'can significantly enhance the quality of the model's responses during the styleless phase.' Without a quantitative or human-judged measure of style neutrality in the first-stage output, the pipeline may be better described as content generation followed by style transfer, and the claimed benefit of decoupling is not established.
[Experiments, 'Human Assessment'] The human evaluation comprises 17 raters and one dialogue per character per method, with no standard deviations, confidence intervals, or significance tests reported. The abstract and the Experiments section use the word 'significantly,' but the data as presented do not support statistical significance. Additionally, for Quality of Conversation among linguistics experts, TTM ranks below Gemini-2.5-pro and Qwen-3-235B-A22B, a result the paper attributes to response length; this should be analyzed rather than only described. Please provide per-item scores, inter-rater agreement, and appropriate statistical tests.
[Experiments, 'Linguistic Style'] The adversarial detection study is a useful idea, but the reported comparison (TTM identified 15.48% of the time vs. Qwen-2.5-32B-Instruct identified 24.80% of the time) is presented without any measure of uncertainty or a significance test. Because the base model in this study is Qwen-2.5-32B-Instruct rather than the Qwen-3-32B used in the main experiments, this result cannot directly support the main claim about the primary system. Clarify whether the observed difference is statistically significant and discuss how it transfers to the main base model.

minor comments (4)

[Figure 3] The label 'Styless Response' in Figure 3 appears to be a typo; it should read 'Styleless Response' consistently with the text and Figure 1.
[Table 1] The table footnote notes that Gemini-2.5-pro employs online grounding, which is a confound in a method comparison. Please state whether any other method had internet or non-text access, and discuss the implications for the comparison.
[Appendix, 'More Discussions'] The first sentence of the 'More Discussions' subsection, 'Though TTM reaches the best performance when againsting with other methods,' contains a typo and overstates the expert Q.C. result; please revise the wording and qualify the claim.
[Sample dialogues in Appendix] In the Dumbledore sample, the user turn contains 'Professer Dumbledore'; if this is an intentional transcription of the user's speech, mark it as such, otherwise correct the spelling.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity; one auxiliary style-detection test is self-referential because it draws distractors from the same historical-utterance database used to produce the response.

other [Experiments, 'Linguistic Style' user study; generation side in Method, 'Linguistic Prompting and Progressive Matching']
"For each question, participants are asked to select the sentence generated by LLMs, from among three examples drawn from the role's historical utterances in books. The historical utterances are selected by performing vector similarity matching between the inserted LLM-generated sentences and the role's historical database."

TTM's style-transfer stage is explicitly conditioned on the same corpus: 'we can use it to retrieve semantically similar utterances from the character's history conversations. These retrieved examples, which are semantically similar to the target input, serve as strong stylistic references to guide the LLM during the linguistic rewriting process.' The detection test then selects its distractors from that same historical-utterance database by vector similarity to the generated sentence. The measured 'style improvement' therefore partly reduces to how faithfully the pipeline copies or paraphrases its own retrieval input: a response that closely follows the retrieved reference will be hard for humans to distinguish from that reference by construction.

full rationale

The paper's central contribution is an empirical pipeline claim, not a closed-form derivation: TTM's three-stage generation is compared against ChatHaruhi, CoSER, GPT-4o, Gemini-2.5-pro, and Qwen-3 through human raters and a GPT-4.1 judge (Table 1), so the headline quality numbers are not fitted from, nor defined in terms of, TTM's extracted personality, memory, or style features. There are no load-bearing same-author citations and no uniqueness theorem imported from the authors' prior work. The main validity weaknesses are non-circular: the Method section concedes that RAG-retrieved content 'may still contain stylistically marked expressions' and that style removal 'may change the original meaning of the sentence,' so the 'styleless' stage is not independently verified, and the appendix concedes that pretraining knowledge of the target character 'can significantly enhance the quality of the model's responses during the styleless phase,' which confounds the decoupling claim. One auxiliary evaluation, the linguistic-style adversarial user study, is mildly self-referential because both the generation's style references and the test's distractor sentences come from the same historical-utterance database, selected by vector similarity; this inflates that particular style-fidelity result without altering the independence of the central human-preference evaluation. Proportional to the evidence, the circularity score is 2.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the pipeline components (personality profile, memory database, style definition) are representations, not independent entities. The free parameters are hand-chosen hyperparameters, not fitted to an objective target. The axioms are domain assumptions about LLM reliability and the separability of dialogue components.

free parameters (6)

chunk_size = 512 tokens
Hand-chosen to balance context and computational cost; directly affects extraction and retrieval quality (Appendix Implementation).
chunk_overlap = 64 tokens
Set to balance coverage and efficiency alongside chunk size.
retrieval_weight_BM25_vs_embedding = [0.5, 0.5]
Fixed equal weights for lexical (BM25) and semantic (Qwen-3-Embedding-0.6B) retrieval in hybrid retrieval of historical utterances.
LLM_as_judge_temperature = 0.2
Used in GPT-4.1 judge; low temperature for stable scoring. Does not affect the method itself but affects evaluation outcomes.
LLM_as_judge_top_p = 0.8
Nucleus sampling parameter for the judge.
number_of_history_utterances_for_style_matching = Not specified
The paper states this number significantly impacts results ('When the number is insufficient, the LLM may overly consider stylistic adaptation...'), implying tuning, but no value or selection procedure is reported.

assumptions (4)

domain assumption Personality, memory, and linguistic style are separable and independently controllable components of dialogue behavior.
The entire method and the paper's title rest on this premise. No experiment directly verifies independence; the paper acknowledges leakage risk in style removal.
domain assumption An LLM can extract accurate personality, background, and style profiles from raw text chunks.
The extraction stage uses Qwen-2.5-32B-Instruct to produce structured profiles; errors would propagate.
domain assumption A 'styleless response' can be generated from personality and background alone, and later style transfer can re-add stylistic markers without altering content.
This is the core mechanism of stage one and stage three. The paper itself cautions that style removal may change meaning for complex or archaic language.
domain assumption Retrieved similar utterances from the character's history provide a valid style reference for rewriting.
Linguistic prompting relies on hybrid retrieval to find stylistically relevant examples; retrieval failures directly degrade the final style.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Test-Time-Matching: Decouple Personality, Memory, and Linguistic Style in LLM-based Role-Playing Language Agent." pith.science (2026). https://pith.science/paper/OQKPG2J4

@misc{pith2026250716799,
  author       = {Pith},
  title        = {Pith review of: Test-Time-Matching: Decouple Personality, Memory, and Linguistic Style in LLM-based Role-Playing Language Agent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OQKPG2J4}},
  note         = {Machine review of arXiv:2507.16799}
}

read the original abstract

The rapid advancement of large language models (LLMs) has enabled role-playing language agents to demonstrate significant potential in various applications. However, relying solely on prompts and contextual inputs often proves insufficient for achieving deep immersion in specific roles, particularly well-known fictional or public figures. On the other hand, fine-tuning-based approaches face limitations due to the challenges associated with data collection and the computational resources required for training, thereby restricting their broader applicability. To address these issues, we propose Test-Time-Matching (TTM), a training-free role-playing framework through test-time scaling and context engineering. TTM uses LLM agents to automatically decouple a character's features into personality, memory, and linguistic style. Our framework involves a structured, three-stage generation pipeline that utilizes these features for controlled role-playing. It achieves high-fidelity role-playing performance, also enables seamless combinations across diverse linguistic styles and even variations in personality and memory. We evaluate our framework through human assessment, and the results demonstrate that our method achieves the outstanding performance in generating expressive and stylistically consistent character dialogues.

Figures

Figures reproduced from arXiv: 2507.16799 by the authors.

**Figure 2.** We present the average ratings from the LLM [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Pipeline overview. TTM first automatically extracts role-relevant information from textual inputs and subsequently [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Memory-checked Response. In the second stage, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: The details of progressive matching and hybrid re [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: TTM achieves a high-fidelity reproduction of a tar [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Visualization of the evaluation results. [PITH_FULL_IMAGE:figures/full_fig_p007_7.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.

RedditPersona: A Modular Framework for Community-Conditioned LLM Adaptation from Reddit
cs.AI 2026-06 unverdicted novelty 6.0 of 10

RedditPersona standardizes community-conditioned LLM adaptation on Reddit via five user-grouping strategies, QLoRA adapters, and a shared evaluation suite, showing identifiability tracks baseline agreement and a consi...

Reference graph

Works this paper leans on

49 extracted references · 34 canonical work pages · cited by 1 Pith paper

[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]

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]

Anuttacon. 2025. Whispers from the Star. https://www.whispersfromthestar.org

work page 2025
[4]

Buongiorno, S.; Klinkert, L.; Zhuang, 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 AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment

work page 2024
[5]

ByteDance. 2024. Maoxiang. https://maoxiangai.com

work page 2024
[6]

Character.AI. 2023. Character AI. https://character.ai

work page 2023
[7]

Chen, J.; Wang, X.; Xu, R.; Yuan, S.; Zhang, Y.; Shi, W.; Xie, J.; Li, S.; Yang, R.; Zhu, T.; et al. 2024. From Persona to Personalization: A Survey on Role-Playing Language Agents. Trans. Mach. Learn. Res.(TMLR)

work page 2024
[8]

Dai, N.; Liang, J.; Qiu, X.; and Huang, X.-J. 2019. Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation. In Annu. Meet. Assoc. Comput. Linguist.(ACL Long)

work page 2019

Show all 49 references

[9]

Dai, Y.; Hu, H.; Wang, L.; Jin, S.; Chen, X.; and Lu, Z. 2025. Mmrole: A comprehensive framework for developing and evaluating multimodal role-playing agents. Int. Conf. Learn. Represent.(ICLR)

2025
[10]

Google. 2025. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities. https://deepmind.google/models/gemini/pro

2025
[11]

Guo, Z.; Xia, L.; Yu, Y.; Ao, T.; and Huang, C. 2024. LightRAG: Simple and Fast Retrieval-Augmented Generation. arXiv preprint arXiv:2410.05779

2024 arXiv
[12]

Han, S.; Chen, L.; Lin, L.-M.; Xu, Z.; and Yu, K. 2024. IBSEN: Director-Actor Agent Collaboration for Controllable and Interactive Drama Script Generation. In Annu. Meet. Assoc. Comput. Linguist.(ACL Long)

2024
[13]

Hu, Z.; Yang, Z.; Liang, X.; Salakhutdinov, R.; and Xing, E. P. 2017. Toward controlled generation of text. In Int. Conf. Mach. Learn.(ICML)

2017
[14]

Huang, L.; Lan, H.; Sun, Z.; Shi, C.; and Bai, T. 2024. Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval. arXiv preprint arXiv:2410.23041

2024 arXiv
[15]

P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al

Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

2024 arXiv
[16]

Lample, G.; Subramanian, S.; Smith, E.; Denoyer, L.; Ranzato, M.; and Boureau, Y.-L. 2019. Multiple-attribute text rewriting. In Int. Conf. Learn. Represent.(ICLR)

2019
[17]

Lee, D.; Tian, Z.; Xue, L.; and Zhang, N. L. 2021. Enhancing content preservation in text style transfer using reverse attention and conditional layer normalization. arXiv preprint arXiv:2108.00449

2021 arXiv
[18]

Li, C.; Leng, Z.; Yan, C.; Shen, J.; Wang, H.; Mi, W.; Fei, Y.; Feng, X.; Yan, S.; Wang, H.; et al. 2023 a . Chatharuhi: Reviving anime character in reality via large language model. arXiv preprint arXiv:2308.09597

2023 arXiv
[19]

Li, J.; Zhang, Z.; Chen, X.; Zhao, D.; and Yan, R. 2023 b . Stylized dialogue generation with feature-guided knowledge augmentation. In Conf. Empirical Methods Nat. Lang. Process.(EMNLP Findings)

2023
[20]

Li, J.; Zhang, Z.; Tu, Q.; Cheng, X.; Zhao, D.; and Yan, R. 2024. Stylechat: Learning recitation-augmented memory in llms for stylized dialogue generation. arXiv preprint arXiv:2403.11439

2024 arXiv
[21]

Lu, K.; Yu, B.; Zhou, C.; and Zhou, J. 2024. Large Language Models are Superpositions of All Characters: Attaining Arbitrary Role-play via Self-Alignment. In Annu. Meet. Assoc. Comput. Linguist.(ACL Long)

2024
[22]

C.; and Lee, H

Lyu, Y.; Luo, T.; Shi, J.; Hollon, T. C.; and Lee, H. 2023. Fine-grained text style transfer with diffusion-based language models. arXiv preprint arXiv:2305.19512

2023 arXiv
[23]

Ma, X.; Gong, Y.; He, P.; Zhao, H.; and Duan, N. 2023. Query rewriting in retrieval-augmented large language models. In Conf. Empirical Methods Nat. Lang. Process.(EMNLP)

2023
[24]

MiniMax. 2023. Talkie. https://www.talkie-ai.com

2023
[25]

OpenAI. 2025. Introducing GPT-4.1 in the API. https://openai.com/index/gpt-4-1

2025
[26]

Pan, L.; Lan, Y.; Li, Y.; and Qian, W. 2024. Unsupervised Text Style Transfer via LLMs and Attention Masking with Multi-way Interactions. arXiv preprint arXiv:2402.13647

2024 arXiv
[27]

E.; and Walker, S

Robertson, S. E.; and Walker, S. 1994. Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In ACM SIGIR Conf. Res. Dev. Inf. Retr.(SIGIR)

1994
[28]

Roy, S.; Shu, R.; Pappas, N.; Mansimov, E.; Zhang, Y.; Mansour, S.; and Roth, D. 2023. Conversation style transfer using few-shot learning. arXiv preprint arXiv:2302.08362

2023 arXiv
[29]

Ryu, J.; Kim, K.; Heo, D.; Song, H.; Oh, C.; and Suh, B. 2025. Cinema Multiverse Lounge: Enhancing Film Appreciation via Multi-Agent Conversations. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems

2025
[30]

Shao, Y.; Li, L.; Dai, J.; and Qiu, X. 2023. Character-LLM: A Trainable Agent for Role-Playing. In Conf. Empirical Methods Nat. Lang. Process.(EMNLP)

2023
[31]

Tao, Z.; Xi, D.; Li, Z.; Tang, L.; and Xu, W. 2024. CAT-LLM: prompting large language models with text style definition for Chinese article-style transfer. arXiv preprint arXiv:2401.05707

2024 arXiv
[32]

Toshevska, M.; and Gievska, S. 2021. A review of text style transfer using deep learning. IEEE Trans. Artif. Intell.(TAI)

2021
[33]

Tu, Q.; Chen, C.; Li, J.; Li, Y.; Shang, S.; Zhao, D.; Wang, R.; and Yan, R. 2023. Characterchat: Learning towards conversational ai with personalized social support. arXiv preprint arXiv:2308.10278

2023 arXiv
[34]

Tu, Q.; Fan, S.; Tian, Z.; and Yan, R. 2024. Charactereval: A chinese benchmark for role-playing conversational agent evaluation. arXiv preprint arXiv:2401.01275

2024 arXiv
[35]

Wang, N.; Que, H.; Liu, J.; Zhou, W.; Wu, Y.; Guo, H.; Gan, R.; Ni, Z.; Yang, J.; Zhang, M.; et al. 2024 a . RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Language Models. In Annu. Meet. Assoc. Comput. Linguist.(ACL Findings)

2024
[36]

Wang, X.; Wang, H.; Zhang, Y.; Yuan, X.; Xu, R.; Huang, J.-t.; Yuan, S.; Guo, H.; Chen, J.; Zhou, S.; et al. 2025. Coser: Coordinating llm-based persona simulation of established roles. In Int. Conf. Mach. Learn.(ICML)

2025
[37]

Wang, X.; Xiao, Y.; Huang, J.-t.; Yuan, S.; Xu, R.; Guo, H.; Tu, Q.; Fei, Y.; Leng, Z.; Wang, W.; et al. 2024 b . InCharacter: Evaluating Personality Fidelity in Role-Playing Agents through Psychological Interviews. In Annu. Meet. Assoc. Comput. Linguist.(ACL Long)

2024
[38]

Weng, L. 2023. LLM Powered Autonomous Agents. https://lilianweng.github.io/posts/2023-06-23-agent

2023
[39]

Xiao, S.; Liu, Z.; Zhang, P.; and Muennighoff, N. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv preprint arXiv:2309.07597

2023 arXiv
[40]

Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025 a . Qwen3 technical report. arXiv preprint arXiv:2505.09388

2025 arXiv
[41]

Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; Lin, H.; Yang, J.; Tu, J.; Zhang, J.; Yang, J.; Yang, J.; Zhou, J.; Lin, J.; Dang, K.; Lu, K.; Bao, K.; Yang, K.; Yu, L.; Li, M.; Xue, M.; Zhang, P.; Zhu, Q.; Men, R.; Lin, R.; Li, ...

2025 arXiv
[42]

Yang, B.; Liu, D.; Xiao, C.; Zhao, K.; Tang, C.; Li, C.; Yuan, L.; Yang, G.; Huang, L.; and Lin, C. 2024. Crafting customisable characters with llms: Introducing simschat, a persona-driven role-playing agent framework. arXiv preprint arXiv:2406.17962

2024
[43]

Yu, X.; Luo, T.; Wei, Y.; Lei, F.; Huang, Y.; Peng, H.; and Zhu, L. 2024. Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent. In Conf. Empirical Methods Nat. Lang. Process.(EMNLP)

2024
[44]

Zhang, B.; Huang, Y.; Cui, W.; and Zhang, H. 2024 a . Thinking Before Speaking: A Role-playing Model with Mindset. arXiv preprint arXiv:2409.13752

2024 arXiv
[45]

Zhang, C.; Cai, H.; Wu, Y.; Hou, L.; Abdul-Mageed, M.; et al. 2024 b . Distilling text style transfer with self-explanation from LLMs. arXiv preprint arXiv:2403.01106

2024 arXiv
[46]

Zhang, Y.; Li, M.; Long, D.; Zhang, X.; Lin, H.; Yang, B.; Xie, P.; Yang, A.; Liu, D.; Lin, J.; Huang, F.; and Zhou, J. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv preprint arXiv:2506.05176

2025 arXiv
[47]

Zhou, J.; Chen, Z.; Wan, D.; Wen, B.; Song, Y.; Yu, J.; Huang, Y.; Ke, P.; Bi, G.; Peng, L.; et al. 2024. CharacterGLM: Customizing Social Characters with Large Language Models. In Conf. Empirical Methods Nat. Lang. Process.(EMNLP Industry Track)

2024
[48]

Zhou, Y.; Su, Y.; Sun, Y.; Wang, S.; Wang, T.; He, R.; Zhang, Y.; Liang, S.; Liu, X.; Ma, Y.; et al. 2025. In-depth Analysis of Graph-based RAG in a Unified Framework. arXiv preprint arXiv:2503.04338

2025 arXiv
[49]

Zhu, X.; Guan, J.; Huang, M.; and Liu, J. 2023. StoryTrans: Non-Parallel Story Author-Style Transfer with Discourse Representations and Content Enhancing. In Annu. Meet. Assoc. Comput. Linguist.(ACL Long)

2023

Pith tools

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

[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]

Anuttacon. 2025. Whispers from the Star. https://www.whispersfromthestar.org

work page 2025

[4] [4]

Buongiorno, S.; Klinkert, L.; Zhuang, 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 AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment

work page 2024

[5] [5]

ByteDance. 2024. Maoxiang. https://maoxiangai.com

work page 2024

[6] [6]

Character.AI. 2023. Character AI. https://character.ai

work page 2023

[7] [7]

Chen, J.; Wang, X.; Xu, R.; Yuan, S.; Zhang, Y.; Shi, W.; Xie, J.; Li, S.; Yang, R.; Zhu, T.; et al. 2024. From Persona to Personalization: A Survey on Role-Playing Language Agents. Trans. Mach. Learn. Res.(TMLR)

work page 2024

[8] [8]

Dai, N.; Liang, J.; Qiu, X.; and Huang, X.-J. 2019. Style Transformer: Unpaired Text Style Transfer without Disentangled Latent Representation. In Annu. Meet. Assoc. Comput. Linguist.(ACL Long)

work page 2019