Pith. sign in

REVIEW 3 major objections 5 minor 34 references

RoleRAG: Enhancing LLM Role-Playing via Graph Guided Retrieval

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read RoleRAG, a retrieval-based framework that merges duplicate entity names and rejects out-of-scope questions, improves role-playing fidelity across general and role-specific LLMs without fine-tuning.

desk verdict Nice system, weak evaluation: the paper's own admission of non-discriminative LLM scores undercuts the abstract's claim, but the entity normalization and boundary-aware retrieval are worth refereeing. read the letter →

arxiv 2505.18541 v1 pith:XZF5GGA4 submitted 2025-05-24 cs.AI

Yongjie Wang , Jonathan Leung , Zhiqi Shen This is my paper
classification cs.AI
keywords RoleRAGretrieval-augmentedgenerationknowledgegraphentitynormalizationrole-playinghallucinationboundary-awareretrievalcharactermodeling
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

RoleRAG is a retrieval-based framework that lets large language models play characters without being fine-tuned for each role. The paper argues that role-play failures stem from two causes: character names that vary across a story, so retrieval misses relevant knowledge, and models that do not know where a character's knowledge ends, so they answer with out-of-universe or fabricated content. RoleRAG builds a structured knowledge graph whose nodes are deduplicated by a semantic entity normalization algorithm, then uses a boundary-aware retriever to fetch character-specific details while rejecting out-of-scope questions. Experiments on three role-playing benchmarks show that this setup improves knowledge exposure, reduces hallucinated responses, and increases refusal of unknown questions for both general-purpose and role-specific language models, sometimes allowing a smaller model to beat a larger one.

What carries the argument

The load-bearing mechanism is a two-part retrieval architecture on a structured character knowledge graph. First, a semantic entity normalization algorithm links name variants by embedding similarity followed by an LLM confirmation step, clusters them by connected components, and assigns a single canonical name to each cluster, so the graph has one node per real-world entity. Second, a boundary-aware retriever processes a user query via a hypothetical-context step (inspired by HyDE), asks an LLM to list the entities in the query with their relevance to the character and their specificity level, and then applies three strategies: reject out-of-scope entities, retrieve detailed entity and relation descriptions for specific entities, and pull type-filtered one-hop neighbors for general entities. This design simultaneously expands the supply of character-specific facts and the model's ability to recognize when a question should be refused.

What would settle it

Have a panel of human evaluators who are deeply familiar with the characters score the same set of role-specific and out-of-scope questions for factual accuracy and fabrication, without knowing which system produced each answer; if RoleRAG does not show a clear reduction in fabricated details and a higher rate of appropriate refusal compared with the plain RAG baseline, the framework's claimed advantage is not real.

Watch

Extended reading notes

Core claim

The central claim is that a fine-tuning-free pipeline—entity extraction, semantic normalization of duplicate names, knowledge graph construction, and boundary-aware retrieval—substantially improves a model's alignment with a character's actual knowledge and reduces role-specific hallucinations. The entity normalization step merges names like 'Anakin Skywalker' and 'Darth Vader' into one canonical entity, so the retrieval layer sees all related facts regardless of the name variant. The retriever then classifies each entity in a user query as specific, general, or out of scope: specific entities are looked up in the knowledge graph, general entities are answered from the character's one-hop neighborhood, and out-of-scope questions are explicitly rejected with a rationale. The authors claim that this structured access to character knowledge is what lets role-playing models stay in character and avoid fabricating details, even for minor characters that the base model knows little about.

Load-bearing premise

The measured improvements depend on evaluation scores that an LLM judge tends to assign in a compressed high range, leaving little room for observable differences between methods.

Editorial extensions

If this is right

  • RoleRAG removes the need for per-character fine-tuning, so a new character can be added by simply building a knowledge graph for that character's source material.
  • Smaller open-language models can match or exceed larger models on role-playing fidelity when given RoleRAG's retrieved context.
  • The boundary-aware rejection mechanism teaches models to refuse out-of-universe questions, a key form of self-awareness in role-playing.
  • The entity normalization step can serve as a general preprocessing tool for any knowledge graph built from noisy or ambiguous text.

Reading between the lines

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

  • The same boundary-aware retrieval pattern could be transferred to any agent that must respect a knowledge boundary, such as a customer-support bot that should not answer policy questions from the wrong era or department.
  • Because the paper's evaluation relies on a judge LLM that gives compressed high scores, the true effect size of RoleRAG might be smaller than the tables suggest; a more discriminative human-verified metric could change the ranking.
  • Combining RoleRAG with a small amount of fine-tuning on a few high-value characters might give the best of both worlds: strong general retrieval plus a stable persona for frequently used roles.
  • The entity normalization method could be evaluated independently on a generic entity-linking benchmark to test whether it scales to much larger knowledge bases.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes RoleRAG, a retrieval-augmented framework for LLM role-playing that combines an entity normalization module (which merges alias names referring to the same character) with a boundary-aware retrieval module (which retrieves character-specific entity details for in-scope queries and encourages rejection of out-of-scope questions). The framework is evaluated on three role-playing datasets (Harry Potter, RoleBench-zh, CharacterLLM) with six base models, comparing against vanilla prompting, chunk-based RAG, a character-profile baseline, and GraphRAG. The reported metrics are knowledge exposure (KE), knowledge hallucination (KH), and unknown-question rejection (UQR), obtained from GPT-4o as a judge with human verification on most datasets. The authors claim consistent improvements in KE and UQR and reductions in KH across models and datasets.

Significance. If the reported gains are reliable, RoleRAG is a plausible lightweight, fine-tuning-free method for improving character fidelity and reducing out-of-character fabrication in role-playing LLMs, and the entity normalization algorithm (with its ≈|N|/k reduction in LLM calls) is a useful contribution for graph-based character indexing. The paper is commendably explicit about the limitations of its evaluation: Section 7 admits that LLM judges assign over-confident scores that are insufficiently discriminative, and Section 5.1 acknowledges that high baselines leave limited room for observable improvement. These admissions are important because the central claim rests on small average score differences, and the paper provides no statistical significance tests or variance estimates. The framework itself is coherent and the ablation study is structured as a 2×2 design, which is a strength. However, the empirical evidence as presented is not yet strong enough to establish the claimed advantages over baselines.

major comments (3)
  1. [Section 5.1, Section 7, Table 1] The central claim that RoleRAG improves character-knowledge alignment and reduces hallucination rests entirely on the GPT-4o judge scores in Table 1, yet the paper itself states in Section 7 that LLM judges 'assign over-confident scores ... render the scores insufficiently discriminative' and in Section 5.1 that high baseline scores 'leave limited room for observable improvement.' The reported differences are small on a 1–10 scale (e.g., KE gains of 0.1–0.3 for Mistral-Small, Llama3.1, and Qwen2.5 on Harry Potter; KH reductions of 0.1–0.4), and no significance tests, confidence intervals, or effect sizes are reported. Without evidence that these differences exceed judge noise, the improvements cannot be distinguished from artifacts of the evaluation procedure. Please provide per-item score distributions, pairwise significance tests (e.g., bootstrap or Wilcoxon signed-rank), and ideally a calibrated judge or an independently human-scored subset for all datasets.
  2. [Appendix C] The human verification procedure is not independent of the LLM judge. Appendix C states that the human evaluator receives the LLM's analysis and 'can adjust the score if they feel that it is not correct,' and the paper also says these LLM analyses can 'mislead human evaluators' (Section 7). This creates a circularity risk in the purported human validation: the human scores may be anchored to the LLM's potentially biased reasoning. Please specify the number of human-evaluated responses, report inter-annotator agreement, and either collect human scores blind to the LLM analysis or show that the human adjustments are not systematically correlated with the LLM's own scoring errors.
  3. [Table 1, footnote ‡] The CharacterLLM dataset, which contains 859 questions (the largest dataset in the study), is evaluated by averaging only two trials of GPT-4o, with no human verification. The general statement in the table caption that 'Human evaluators are recruited to verify and correct GPT-4o's score' is therefore inaccurate for this dataset. Since several reported RoleRAG advantages on CharacterLLM are very small (e.g., KE gain over GraphRAG of 0.09 for Qwen2.5 and 0.002 for Llama3.3), the absence of human verification on the largest dataset further weakens the empirical support for the central claim. Please either human-evaluate a random subset with a pre-specified sample size, or explicitly qualify that CharacterLLM results are LLM-judge-only and discuss the potential bias this introduces.
minor comments (5)
  1. [Table 1 vs. Section 4.1] The baseline called 'Character profile' in Section 4.1 is labeled 'User profile' in Table 1; please align the terminology throughout.
  2. [Table 5] The header says characters are listed 'from highest to lowest frequency,' but the text elsewhere refers to Voldemort as 'Voldemort' and the table inconsistently uses 'V oldemort' (with a space). Please fix the typo and confirm the ordering criterion.
  3. [Table 2] The rows 'Without Local search' and 'With Local search' are ambiguous; they should read 'Without entity normalization, Local search' and 'With entity normalization, Local search,' respectively, to make the factorial design clear.
  4. [Appendix C] The evaluation section says the judge's temperature is set to 0.2, but no justification is given for this specific value; cite prior work or report sensitivity to temperature, since the judge is a core component of the reported results.
  5. [Section 7] The limitations paragraph beginning 'Another limitation of our work' appears after two separate limitations and is followed by further limitations; please restructure to enumerate all limitations (e.g., as a numbered list) for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: RoleRAG's components are evaluated by external performance metrics; the admitted score saturation is a validity concern, not a circular reduction.

full rationale

The paper derives no mathematical result or prediction that reduces to its own inputs by construction. The entity normalization module and boundary-aware retrieval strategy are algorithmic components whose effectiveness is assessed through downstream role-playing performance metrics (knowledge exposure, hallucination, unknown-question rejection) on held-out datasets. No fitted parameter is renamed as a prediction; no equation defines the target metric in terms of the retrieval output. The only self-referential element is the use of the GPT model family for both pipeline construction (GPT-4o mini) and evaluation (GPT-4o as judge), which is a methodological bias concern rather than a logical circularity. The paper's own admission that LLM judges assign over-confident scores, leaving 'limited room for observable improvement' (Section 5.1) and that scores are 'insufficiently discriminative' (Section 7), is a threat to the validity of the empirical claims, but it does not make the derivation circular. Since the central claim is an empirical comparison supported by ablations and external baselines, the paper is self-contained against non-circular evaluation.

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

The ledger lists the main hand-chosen parameters of the pipeline and the domain assumptions that the system takes for granted. None of these are fitted to the target result, but they are load-bearing for the empirical claim.

free parameters (4)
  • chunk_size = 600 tokens
    Chosen by hand for splitting character profiles into processing chunks (Section 4.4).
  • chunk_overlap = 100 tokens
    Chosen by hand for chunk segmentation (Section 4.4).
  • k (top-k retrieval) = not specified in paper
    Used in entity normalization (Algorithm 1) and retrieval; the value is not reported, so the system cannot be exactly reproduced.
  • judge_temperature = 0.2
    Set for GPT-4o judge to reduce creativity (Appendix C). This affects evaluation but not the method itself.
assumptions (4)
  • domain assumption LLM-based entity and relation extraction from chunks is accurate enough for graph construction
    The entire knowledge graph is built from LLM-extracted triples (Section 3.1). Errors propagate to retrieval and response quality.
  • domain assumption Embedding similarity is a valid proxy for semantic entity equivalence in the normalization step
    Algorithm 1 relies on top-k embedded entities to propose candidate merges (Section 3.2). If embeddings miss relevant variants, duplicates are never merged.
  • domain assumption The constructed knowledge graph contains sufficient knowledge to answer the test questions
    Retrieval draws only from the graph built from Wikipedia, Baidu Baike, and novels (Section 4.1). Missing facts cannot be recovered.
  • domain assumption The LLM's relevance and specificity judgment correctly identifies out-of-scope entities
    Boundary-aware rejection depends on the LLM's ability to judge whether an entity is within the character's knowledge (Section 3.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of RoleRAG: Enhancing LLM Role-Playing via Graph Guided Retrieval." pith.science (2026). https://pith.science/paper/XZF5GGA4

@misc{pith2026250518541,
  author       = {Pith},
  title        = {Pith review of: RoleRAG: Enhancing LLM Role-Playing via Graph Guided Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZF5GGA4}},
  note         = {Machine review of arXiv:2505.18541}
}
read the original abstract

Large Language Models (LLMs) have shown promise in character imitation, enabling immersive and engaging conversations. However, they often generate content that is irrelevant or inconsistent with a character's background. We attribute these failures to: (1) the inability to accurately recall character-specific knowledge due to entity ambiguity, and (2) a lack of awareness of the character's cognitive boundaries. To address these issues, we propose RoleRAG, a retrieval-based framework that integrates efficient entity disambiguation for knowledge indexing with a boundary-aware retriever for extracting contextually appropriate information from a structured knowledge graph. Experiments on role-playing benchmarks show that RoleRAG's calibrated retrieval helps both general-purpose and role-specific LLMs better align with character knowledge and reduce hallucinated responses.

Figures

Figures reproduced from arXiv: 2505.18541 by the authors.

Figure 1
Figure 1. This figure illustrates that LLMs perform [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Workflow of our proposed RoleRAG [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of evaluation metrics. We encourage LLMs to exhibit more personal traits, minimize fabricated [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Experiments of out-of-scope questions in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Word cloud for responses generated by GPT [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Word cloud for responses generated by GPT [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Use case of retrieval step in our RoleRAG. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: The prompt used for generating knowledge exposure scores. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: The prompt used for generating knowledge hallucination scores. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: The prompt used for generating unknown question rejection scores. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: The prompt used for generating responses on the Harry Potter dataset. We use the colon character (":") [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 5 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  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]

    Yanqi Dai, Huanran Hu, Lei Wang, Shengjie Jin, Xu Chen, and Zhiwu Lu. 2024. https://arxiv.org/abs/2408.04203 Mmrole: A comprehensive framework for developing and evaluating multimodal role-playing agents . arXiv preprint arXiv:2408.04203

    arXiv 2024

  4. [4]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    arXiv 2024

  5. [5]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, and Jonathan Larson. 2024. https://arxiv.org/abs/2404.16130 From Local to Global: A Graph RAG Approach to Query-Focused Summarization . pages 1--15

    arXiv 2024

  6. [6]

    Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. 2023. https://doi.org/10.18653/v1/2023.acl-long.99 Precise zero-shot dense retrieval without relevance labels . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1762--1777, Toronto, Canada. Association for Computational Linguistics

    doi:10.18653/v1/2023.acl-long.99 2023

  7. [7]

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. Lightrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779

    arXiv 2024

  8. [8]

    Hang Jiang, Xiajie Zhang, Xubo Cao, Cynthia Breazeal, Deb Roy, and Jad Kabbara. 2023. https://arxiv.org/abs/2305.02547 Personallm: Investigating the ability of large language models to express personality traits . arXiv preprint arXiv:2305.02547

    arXiv 2023

Show all 34 references
  1. [9]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

    2020

  2. [10]

    Cheng Li, Ziang Leng, Chenxi Yan, Junyi Shen, Hao Wang, Weishi Mi, Yaying Fei, Xiaoyang Feng, Song Yan, HaoSheng Wang, et al. 2023. Chatharuhi: Reviving anime character in reality via large language model. arXiv preprint arXiv:2308.09597

    2023 arXiv

  3. [11]

    Zhuohang Li, Jiaxin Zhang, Chao Yan, Kamalika Das, Sricharan Kumar, Murat Kantarcioglu, and Bradley A. Malin. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.353 Do you know what you are talking about? characterizing query-knowledge relevance for reliable retrieval augmented...

    2024 doi

  4. [12]

    Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022. https://doi.org/10.18653/v1/2022.deelio-1.10 What makes good in-context examples for GPT -3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extr...

    2022 doi

  5. [13]

    Jiongnan Liu, Yutao Zhu, Shuting Wang, Xiaochi Wei, Erxue Min, Yu Lu, Shuaiqiang Wang, Dawei Yin, and Zhicheng Dou. 2024. https://arxiv.org/abs/2409.11901 LLMs + Persona-Plug = Personalized LLMs

    2024 arXiv

  6. [14]

    Keming Lu, Bowen Yu, Chang Zhou, and Jingren Zhou. 2024. https://doi.org/10.18653/v1/2024.acl-long.423 Large language models are superpositions of all characters: Attaining arbitrary role-play via self-alignment . In Proceedings of the 62nd Annual Meeting of the Association fo...

    2024 doi

  7. [15]

    Mistral. 2025. https://mistral.ai/en/news/mistral-small-3 Mistral small 3

    2025

  8. [16]

    OpenAI. 2024. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence Gpt-4o mini: Advancing cost-efficient intelligence

    2024

  9. [17]

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22

    2023

  10. [18]

    Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. 2024. https://doi.org/10.18653/v1/2024.acl-long.399 L a MP : When large language models meet personalization . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volu...

    2024 doi

  11. [19]

    Bhaskarjit Sarmah, Dhagash Mehta, Benika Hall, Rohan Rao, Sunil Patel, and Stefano Pasquali. 2024. https://doi.org/10.1145/3677052.3698671 Hybridrag: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction . In Proceedings of...

    2024

  12. [20]

    Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. 2023. https://aclanthology.org/2023.emnlp-main.814 Character- LLM : A trainable agent for role-playing . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13153--13187, Singapore...

    2023

  13. [21]

    Chenhui Shen, Liying Cheng, Xuan-Phi Nguyen, Yang You, and Lidong Bing. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.278 Large language models are not yet human-level evaluators for abstractive summarization . In Findings of the Association for Computational Linguisti...

    2023 doi

  14. [22]

    Meiling Tao, Liang Xuechen, Tianyu Shi, Lei Yu, and Yiting Xie. 2024. https://aclanthology.org/2024.personalize-1.1 R ole C raft- GLM : Advancing personalized role-playing in large language models . In Proceedings of the 1st Workshop on Personalization of Generative AI Systems...

    2024

  15. [23]

    Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, Wei-Lin Chen, Chao-Wei Huang, Yu Meng, and Yun-Nung Chen. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.969 Two tales of persona in LLM s: A survey of role-playing and personalization . In Findings of the Association for Com...

    2024 doi

  16. [24]

    Quan Tu, Shilong Fan, Zihang Tian, Tianhao Shen, Shuo Shang, Xin Gao, and Rui Yan. 2024. https://doi.org/10.18653/v1/2024.acl-long.638 C haracter E val: A C hinese benchmark for role-playing conversational agent evaluation . In Proceedings of the 62nd Annual Meeting of the Ass...

    2024 doi

  17. [25]

    Noah Wang, Z.y. Peng, Haoran Que, Jiaheng Liu, Wangchunshu Zhou, Yuhan Wu, Hongcheng Guo, Ruitong Gan, Zehao Ni, Jian Yang, Man Zhang, Zhaoxiang Zhang, Wanli Ouyang, Ke Xu, Wenhao Huang, Jie Fu, and Junran Peng. 2024 a . https://doi.org/10.18653/v1/2024.findings-acl.878 R ole ...

    2024 doi

  18. [26]

    Xintao Wang, Yunze Xiao, Jen-tse Huang, Siyu Yuan, Rui Xu, Haoran Guo, Quan Tu, Yaying Fei, Ziang Leng, Wei Wang, Jiangjie Chen, Cheng Li, and Yanghua Xiao. 2024 b . https://doi.org/10.18653/v1/2024.acl-long.102 I n C haracter: Evaluating personality fidelity in role-playing a...

    2024 doi

  19. [27]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  20. [28]

    Nathaniel Weir, Ryan Thomas, Randolph d ' Amore, Kellie Hill, Benjamin Van Durme, and Harsh Jhamtani. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.520 Ontologically faithful generation of non-player character dialogues . In Proceedings of the 2024 Conference on Empirical ...

    2024 doi

  21. [29]

    Junde Wu, Jiayuan Zhu, and Yunli Qi. 2024. https://arxiv.org/pdf/2408.04187 Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation . arXiv preprint arXiv:2408.04187

    2024 arXiv

  22. [30]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  23. [31]

    Gonzalez, and Ion Stoica

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. https://openreview.net/forum?id=uccHPGDlao Judging LLM -as-a-judge with MT -bench and chatbot ...

    2023

  24. [32]

    Hanxun Zhong, Zhicheng Dou, Yutao Zhu, Hongjin Qian, and Ji-Rong Wen. 2022. https://doi.org/10.18653/v1/2022.naacl-main.426 Less is more: Learning to refine dialogue history for personalized dialogue generation . In Proceedings of the 2022 Conference of the North American Chap...

    2022 doi

  25. [33]

    Jinfeng Zhou, Zhuang Chen, Dazhen Wan, Bosi Wen, Yi Song, Jifan Yu, Yongkang Huang, Pei Ke, Guanqun Bi, Libiao Peng, JiaMing Yang, Xiyao Xiao, Sahand Sabour, Xiaohan Zhang, Wenjing Hou, Yijia Zhang, Yuxiao Dong, Hongning Wang, Jie Tang, and Minlie Huang. 2024. https://doi.org/...

    2024 doi

  26. [34]

    Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. 2023. https://openreview.net/forum?id=pV1xV2RK6I Tool QA : A dataset for LLM question answering with external tools . In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.