Pith. sign in

REVIEW 3 major objections 3 minor 42 references

DialogueForge: LLM Simulation of Human-Chatbot Dialogue

T0 review · 3 major / 3 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 DialogueForge shows that synthetic multi-turn human-chatbot dialogues, seeded by a single real utterance and generated by alternating LLM agents, pass LLM judges as human most of the time, and that fine-tuning lets small open models…

desk verdict A solid, honest BotChat derivative with a useful LoRA fine-tuning result, but the headline 'small models catch up' claim relies entirely on GPT-4o judging and lacks a human or alternative-judge check on the fine-tuned models. read the letter →

arxiv 2507.15752 v1 pith:JG4V4JGM submitted 2025-07-21 cs.CL cs.AI

Ruizhe Zhu , Hao Zhu , Yaxuan Li , Syang Zhou , Shijing Cai , Malgorzata Lazuka , Elliott Ash This is my paper · ORCID
classification cs.CLcs.AI
keywords DialogueForgesyntheticgenerationhuman-chatbotsimulationLLM-as-a-judgeUniEvalGTEvalLoRAfine-tuningmulti-turn
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

DialogueForge claims that a cheap, scalable pipeline can synthesize multi-turn human-chatbot conversations that LLM judges often cannot tell apart from real ones. The framework starts from a single real first user utterance, infers the conversation's topic, and then lets one LLM play the human "inquirer" and another play the chatbot "responder." On the OASST1 6-turn benchmark, GPT-4o and GPT-4o mini exceed 90 percent UniEval passing rates, slightly above real conversations, and LoRA fine-tuning lifts small open models like Llama-3.1-8B to GTEval indistinguishability scores comparable to or better than GPT-4o in some settings. If this holds, researchers could replace expensive crowdsourced dialogue collection with synthetic data, and resource-limited teams could use fine-tuned small models instead of large proprietary APIs.

What carries the argument

The load-bearing mechanism is a two-agent generation loop: an "inquirer" LLM simulates the human user and a "responder" LLM, fixed as GPT-4o mini, simulates the chatbot, with the full conversation history passed between them each turn. Each dialogue is seeded with the first user utterance and an inferred topic from a real human-chatbot exchange, and generation stops at a 6- or 12-turn cap or when the inquirer signals conclusion. Evaluation uses two LLM-as-judge protocols adopted from [12]: UniEval, where the judge says whether a dialogue involves AI and locates the first AI utterance, and GTEval, where the judge compares a generated dialogue against a real one. Fine-tuning applies LoRA to small models on a curated cross-dataset corpus, so a model tested on OASST1 is trained on Chatbot Arena dialogues and vice versa.

What would settle it

Have human raters classify a mixed set of real and DialogueForge-generated conversations: if people spot the generated dialogues easily while the LLM judges pass them at 90 percent, the central claim is an artifact of the judge; the same comparison with GPT-4o-generated dialogues rated by GPT-4o versus by a blind human panel would settle whether same-model favoritism inflates the scores.

Watch

Extended reading notes

Core claim

The paper's central claim is that DialogueForge produces synthetic human-chatbot dialogues that are frequently judged human-like, and that this capability is not reserved for large proprietary models. Concretely, the authors report that GPT-4o and GPT-4o mini pass the UniEval "no AI involved" test on more than 90 percent of 6-turn OASST1 dialogues, slightly outscoring real human-chatbot conversations, while fine-tuned Llama-3.1-8B reaches GTEval indistinguishability scores comparable to or better than GPT-4o on some settings. They further report that all models lose realism as dialogues grow from 6 to 12 turns, and that alternative judge LLMs such as Claude 3.7 and Gemini 2.0 Flash yield largely consistent rankings, which they take as evidence against same-model judge bias.

Load-bearing premise

The load-bearing premise is that UniEval and GTEval scores from LLM judges, with GPT-4o as the default judge, actually capture human-likeness; if those judgments do not match human judgments of naturalness, the rankings, the fine-tuning improvements, and the indistinguishability claims all lose support.

Editorial extensions

If this is right

  • If the claims hold, human-chatbot dialogue data for training or evaluation can be produced at scale from a single real seed utterance instead of crowdsourced collection.
  • Fine-tuned small open models become a viable substitute for large proprietary inquirers, lowering the compute and API cost of generating realistic dialogue data.
  • The reported 6-to-12 turn degradation is a direct warning that synthetic-data pipelines should budget for shorter dialogues or invest in length-robustness methods.
  • Cross-judge consistency in the paper suggests the evaluation protocol transfers across LLM judge choices, making UniEval and GTEval reusable screening tools.
  • Cross-dataset fine-tuning effects imply that data quality, not just size, determines how well a small model learns conversational naturalness.

Reading between the lines

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

  • Because the default judge is GPT-4o, the same model that generates the best-scoring dialogues, the cross-judge checks reduce but do not eliminate the possibility of same-model stylistic favoritism; a human-rating study would settle it.
  • Generated dialogues scoring slightly above real ones on UniEval suggests the judge may reward polished, information-dense phrasing rather than the messiness of genuine human typing, which could inflate apparent indistinguishability.
  • A natural extension would be to use DialogueForge output as training data for a downstream dialogue agent and measure task completion or user satisfaction, since the paper explicitly does not evaluate task success, toxicity, or hallucination.
  • Seeding from only two public datasets limits coverage of interaction styles; testing seeds from other sources, such as customer-support or tutoring logs, would show whether the framework's results generalize.
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 / 3 minor

Summary. DialogueForge is a framework for synthesizing multi-turn human-chatbot dialogues. It seeds generation with the first user utterance from a real human-chatbot exchange, infers a task topic, and then alternates an 'Inquirer' LLM (simulating the human) and a 'Responder' LLM (simulating the chatbot) to produce new dialogues. The paper evaluates several proprietary and open-source LLMs as inquirers using two LLM-as-judge protocols, UniEval and GTEval, and investigates whether LoRA fine-tuning of small models (Llama-3.2-3B, Llama-3.1-8B, Mistral-7B) improves their human-likeness. The main empirical findings are: larger models such as GPT-4o generally achieve higher passing/indistinguishability rates; fine-tuning substantially improves small-model scores, in some cases making them competitive with GPT-4o; and all models degrade as dialogue length grows from 6 to 12 turns. The paper also reports an alternative-judge check using Claude 3.7 and Gemini 2.0 Flash on a subset of models, and releases code, models, and data.

Significance. If the evaluation were externally validated, DialogueForge would be a useful contribution to synthetic dialogue data generation, especially for resource-limited settings: the cross-dataset fine-tuning design, the use of two complementary LLM-as-judge protocols, and the public release of code/data/models are all strengths that make the work reproducible and reusable. However, the central claims—that fine-tuned small models become 'competitive' with GPT-4o and that the performance gap between large and small models is 'substantially reduced'—rest entirely on UniEval/GTEval scores assigned by GPT-4o. The manuscript itself concedes that these metrics 'may not fully reflect the nuanced qualities of natural conversations' (Section 5.2) and defers human validation to future work (Section 6). The alternative-judge analysis is too narrow to rule out judge self-preference or metric invalidity, and no confidence intervals or significance tests are reported. The trends are plausible and consistent with prior work, but the strength of the stated conclusions exceeds what the current evidence can support.

major comments (3)
  1. [§4.1, §4.2.5, §5.2] The central claim that fine-tuning closes the gap between small and large models (Section 4.2.2, Tables 1-2) is established solely through UniEval and GTEval scores assigned by GPT-4o, which is also the generator of the top-scoring dialogues. The alternative-judge experiments in Figures 4 and 5 cover only a subset of base models on the 6-turn OASST1 set and exclude all fine-tuned models (Llama-3.2-3B(FT), Llama-3.1-8B(FT), Mistral-7B(FT)) whose improvements drive the headline result. Section 5.2 acknowledges that these metrics may not capture naturalness, and Section 6 postpones human assessment to future work. Therefore, the evidence does not rule out judge self-preference or metric invalidity, and the claim that fine-tuned small models are 'competitive' with GPT-4o is not yet supported.
  2. [§4.2.3, Tables 1-2] No confidence intervals, standard errors, or significance tests are reported for any of the passing rates or indistinguishability rates, despite finite samples of 413 OASST1 and 1,002 Chatbot Arena conversations. This is especially problematic for the comparison in Section 4.2.2 that fine-tuned Llama-3.1-8B achieves a higher GTEval score than GPT-4o; without variance estimates, this difference could easily arise from sampling noise. The paper should provide error bars, bootstrap intervals, or significance tests for the key comparisons, particularly those involving fine-tuned models.
  3. [§4.2.5] The bias analysis is too narrow to support the conclusion that GPT-4o 'does not exhibit a systematic bias' and 'can be considered a reliable LLM judge.' Figures 4 and 5 use only the 6-turn OASST1 subset, cover at most seven inquirer models (not the full set), and omit the fine-tuned models entirely. Furthermore, the Gemini-2.0-Flash result gives Gemma-2-27B a 'surprisingly high' GTEval rate that differs from the GPT-4o ranking; the text reports this divergence but does not reconcile it. The claimed robustness of the metrics across judge choices (Section 1, contribution 3) is therefore overstated by the current experiments.
minor comments (3)
  1. [§3.3] The section first states that fine-tuning uses 'a carefully curated subset of our synthetic dialogue corpus,' but the following paragraph says the fine-tuning data was 'constructed by extracting and formatting samples from the same dialogue corpora as used for seed prompt extraction: OASST1 and Chatbot Arena,' which are real human-chatbot datasets. Please clarify whether the fine-tuning corpus is synthetic, real, or mixed; this is important for reproducibility.
  2. [Figures 2 and 3] The captions of Figures 2 and 3 both say 'of OASST1 dialogues,' but panels (c) and (d) in each figure present Chatbot Arena results. Please correct the captions to match the datasets shown in each panel.
  3. [§3.1.1] The procedure for 'identify[ing] the central topic' of the seed dialogue, which the framework uses to guide generation, is not described in any detail. Please specify how the topic/task objective is inferred (e.g., a separate LLM call, a parsing heuristic), and mention any validation of this inference step.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: evaluation metrics are adopted from external work, fine-tuning uses cross-dataset real dialogues, and the central claims are empirical measurements rather than constructions.

full rationale

The derivation chain is not circular. The human-likeness metrics (UniEval and GTEval) are adopted from BotChat, an external prior work, not defined by the present paper's own outputs. The fact that GPT-4o serves both as a generator and as the default judge is a potential self-preference threat, but the paper partially addresses it in Section 4.2.5 by re-running a subset of evaluations with Claude 3.7 and Gemini 2.0 Flash, and it explicitly acknowledges the limitation that these metrics may not capture all nuances of natural conversation (Section 5.2). The fine-tuning results are not fitted into the evaluation labels: models are fine-tuned on real dialogues from one dataset and evaluated on generated dialogues seeded from the other dataset, a cross-dataset design. No parameter is fitted and then renamed as a prediction, no uniqueness theorem is imported from the authors' own prior work, no ansatz is smuggled in via self-citation, and no known result is merely renamed. The reviewer's concern that GPT-4o's judgments may not track human judgments is an external-validity caveat, not an internal circularity: the paper's claims would remain meaningful even if the judge were biased, though they would be less well supported. Accordingly, no circular step can be exhibited by quoting equations or definitions that reduce the conclusions to their inputs.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new theoretical entities or parameters fitted to data. The central claims rest on domain assumptions about the validity of LLM-as-judge metrics, the representativeness of the seed datasets, and the sufficiency of first-utterance seeding. The only hand-chosen numerical parameters are the turn limits (6 and 12), which directly shape the length-degradation analysis.

free parameters (1)
  • Maximum dialogue turns = 6 and 12
    Chosen by hand to study length effects; the observed degradation over length is a central finding, so the choice of limits shapes the reported trends. Not fitted to data.
assumptions (3)
  • domain assumption UniEval and GTEval LLM judgments are valid proxies for human-like dialogue quality.
    Invoked in Section 4.1 where these metrics are adopted; all conclusions about which models are more 'human-like' depend on this.
  • domain assumption The filtered OASST1 (413 conversations) and Chatbot Arena (1,002 conversations) subsets are representative of human-chatbot interactions.
    Section 3.2.1 states these are the seed datasets without explaining the filtering criteria; if they are biased, the generated dialogue distribution is also biased.
  • domain assumption Using only the first user utterance and an inferred topic is sufficient to generate coherent, task-tailored dialogues.
    Section 3.1.1 defines the seed prompt as the first utterance plus the inferred topic; the paper's own failure cases (topic deviation, refusal) show this assumption is not always satisfied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DialogueForge: LLM Simulation of Human-Chatbot Dialogue." pith.science (2026). https://pith.science/paper/JG4V4JGM

@misc{pith2026250715752,
  author       = {Pith},
  title        = {Pith review of: DialogueForge: LLM Simulation of Human-Chatbot Dialogue},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JG4V4JGM}},
  note         = {Machine review of arXiv:2507.15752}
}
read the original abstract

Collecting human-chatbot dialogues typically demands substantial manual effort and is time-consuming, which limits and poses challenges for research on conversational AI. In this work, we propose DialogueForge - a framework for generating AI-simulated conversations in human-chatbot style. To initialize each generated conversation, DialogueForge uses seed prompts extracted from real human-chatbot interactions. We test a variety of LLMs to simulate the human chatbot user, ranging from state-of-the-art proprietary models to small-scale open-source LLMs, and generate multi-turn dialogues tailored to specific tasks. In addition, we explore fine-tuning techniques to enhance the ability of smaller models to produce indistinguishable human-like dialogues. We evaluate the quality of the simulated conversations and compare different models using the UniEval and GTEval evaluation protocols. Our experiments show that large proprietary models (e.g., GPT-4o) generally outperform others in generating more realistic dialogues, while smaller open-source models (e.g., Llama, Mistral) offer promising performance with greater customization. We demonstrate that the performance of smaller models can be significantly improved by employing supervised fine-tuning techniques. Nevertheless, maintaining coherent and natural long-form human-like dialogues remains a common challenge across all models.

Figures

Figures reproduced from arXiv: 2507.15752 by the authors.

Figure 1
Figure 1. Flow of the dialogue generation and evaluation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. UniEval passing rates (No AI-Involved) of OASST1 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. GTEval indistinguishability rates (with real data) of [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: UniEval passing rate (no AI involved) of the OASST1 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: GTEval indistinguishability rate (compared with [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 11 canonical work pages

  1. [1]

    Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Mérouane Debbah, Étienne Goffinet, Daniel Hesslow, Julien Launay, Quentin Malartic, Daniele Mazzotta, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. 2023. The Falcon Series of Open Language Models. arXiv:2311.16867 [cs.CL] https://arxiv.org/abs/2311.16867

    arXiv 2023

  2. [2]

    Anthropic. 2025. Claude 3.7 Sonnet and Claude Code. https://www.anthropic. com/news/claude-3-7-sonnet/

    work page 2025

  3. [3]

    Ge Bai, Jie Liu, Xingyuan Bu, Yancheng He, Jiaheng Liu, Zhanhui Zhou, Zhuoran Lin, Wenbo Su, Tiezheng Ge, Bo Zheng, and Wanli Ouyang. 2024. MT-Bench-101: A Fine-Grained Benchmark for Evaluating Large Language Models in Multi- Turn Dialogues. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) ...

    doi:10.18653/v 2024

  4. [4]

    Amy Brand, Liz Allen, Micah Altman, Marjorie Hlava, and Jo Scott. 2015. Be- yond authorship: Attribution, contribution, collaboration, and credit. Learned Publishing 28, 2 (2015), 2

    work page 2015

  5. [5]

    Maximillian Chen, Alexandros Papangelis, Chenyang Tao, Seokhwan Kim, Andy Rosenbaum, Yang Liu, Zhou Yu, and Dilek Hakkani-Tur. 2023. PLACES: Prompt- ing Language Models for Social Conversation Synthesis. In Findings of the Asso- ciation for Computational Linguistics: EACL 2023 , Andreas Vlachos and Isabelle Augenstein (Eds.). Association for Computational...

    doi:10.18653/v1/2023.findings-eacl.63 2023

  6. [6]

    Angelopoulos, Tianle Li, Dacheng Li, Banghua Zhu, Hao Zhang, Michael I

    Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios N. Angelopoulos, Tianle Li, Dacheng Li, Banghua Zhu, Hao Zhang, Michael I. Jordan, Joseph E. Gonzalez, and Ion Stoica. 2024. Chatbot arena: an open platform for evaluating LLMs by human preference. In Proceedings of the 41st International Conference on Machine Learning (Vienna, Austria) (ICML’24). JMLR...

    work page 2024

  7. [7]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training Verifiers to Solve Math Word Problems. arXiv:2110.14168 [cs.LG] https://arxiv.org/abs/2110.14168

    arXiv 2021

  8. [8]

    DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, (...), and Zizheng Pan. 2025. DeepSeek-V3 Technical Report. arXiv:2412.19437 [cs.CL] https://arxiv.org/abs/2412.19437

    arXiv 2025

Show all 42 references
  1. [9]

    Yang Deng, Lizi Liao, Liang Chen, Hongru Wang, Wenqiang Lei, and Tat-Seng Chua. 2023. Prompting and Evaluating Large Language Models for Proactive Dialogues: Clarification, Target-guided, and Non-collaboration. In Findings of the Association for Computational Linguistics: EMNL...

    2023 doi

  2. [10]

    Arnout Devos. 2024. Few-shot Learning for Efficient and Effective Machine Learning Model Adaptation. Ph. D. Dissertation. EPFL

    2024

  3. [11]

    Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. 2023. Enhancing Chat Language Models by Scaling High-quality Instructional Conversations. In Proceedings of the 2023 Con- ference on Empirical Methods in Natural Language Proces...

    2023 doi

  4. [12]

    Haodong Duan, Jueqi Wei, Chonghua Wang, Hongwei Liu, Yixiao Fang, Songyang Zhang, Dahua Lin, and Kai Chen. 2023. BotChat: Evaluating LLMs’ Capabilities of Having Multi-Turn Dialogues. arXiv:2310.13650 [cs.CL] https://arxiv.org/abs/ 2310.13650

    2023 arXiv

  5. [13]

    Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. 2023. GPTScore: Evaluate as You Desire. arXiv:2302.04166 [cs.CL] https://arxiv.org/abs/2302.04166

    2023 arXiv

  6. [14]

    Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, Andy Jones, Sam Bowman, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Nelson Elhage, Sheer El-Showk, Stanislav Fort, Zac Hatfi...

    2022 arXiv

  7. [15]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Art...

    2024 arXiv

  8. [16]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Un- derstanding. arXiv:2009.03300 [cs.CY] https://arxiv.org/abs/2009.03300

    2021 arXiv

  9. [17]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations . https: //openreview.net/forum?id=nZeVKeeFYf9

    2022

  10. [18]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thoma...

    2023 arXiv

  11. [19]

    Hyunwoo Kim, Jack Hessel, Liwei Jiang, Peter West, Ximing Lu, Youngjae Yu, Pei Zhou, Ronan Bras, Malihe Alikhani, Gunhee Kim, Maarten Sap, and Yejin Choi. 2023. SODA: Million-scale Dialogue Distillation with Social Commonsense Contextualization. In Proceedings of the 2023 Conf...

    2023

  12. [20]

    Minju Kim, Chaehyeong Kim, Yong Ho Song, Seung-won Hwang, and Jinyoung Yeo. 2022. BotsTalk: Machine-sourced Framework for Automatic Curation of Agentic & GenAI Evaluation Workshop KDD’25, August 4, 2025, Toronto, ON, Canada R. Zhu, H. Zhu, Y. Li, S. Zhou, S. Cai, M. Łazuka, an...

    2022

  13. [21]

    Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotirisr Anagnostidis, Zhi- Rui Tam, Keith Stevens, Abdullah Barhoum, Nguyen Minh Duc, Oliver Stanley, Richárd Nagyfi, Shahul ES, Sameer Suri, David Glushkov, Arnav Dantuluri, An- drew Maguire, Christoph Schuhmann, Huu Nguyen, a...

    2023

  14. [22]

    LangGraph. 2024. LangGraph Overview. https://langchain-ai.github.io/ langgraph/

    2024

  15. [23]

    Young-Jun Lee, Chae-Gyun Lim, Yunsu Choi, Ji-Hui Lm, and Ho-Jin Choi. 2022. PERSONACHATGEN: Generating Personalized Dialogues using GPT-3. In Pro- ceedings of the 1st Workshop on Customized Chat Grounding Persona and Knowl- edge, Heuiseok Lim, Seungryong Kim, Yeonsoo Lee, Stev...

    2022

  16. [24]

    Zekun Li, Wenhu Chen, Shiyang Li, Hong Wang, Jing Qian, and Xifeng Yan. 2022. Controllable Dialogue Simulation with In-context Learning. In Findings of the Association for Computational Linguistics: EMNLP 2022 , Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang (Eds.). Associati...

    2022 doi

  17. [25]

    Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. 2021. Towards Emotional Support Dialog Systems. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Co...

    2021 doi

  18. [26]

    Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, and Ahmed Awadallah. 2023. Orca: Progressive Learning from Complex Explanation Traces of GPT-4. arXiv:2306.02707 [cs.CL] https://arxiv.org/abs/ 2306.02707

    2023 arXiv

  19. [27]

    OpenAI, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Mądry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex Paino,...

  20. [28]

    Heydar Soudani, Roxana Petcu, Evangelos Kanoulas, and Faegheh Hasibi. 2024. Data Augmentation for Conversational AI. In Companion Proceedings of the ACM Web Conference 2024 (Singapore, Singapore) (WWW ’24). Association for Computing Machinery, New York, NY, USA, 1234–1237. doi...

    2024 doi

  21. [29]

    Heydar Soudani, Roxana Petcu, Evangelos Kanoulas, and Faegheh Hasibi

  22. [30]

    Hovhannes Tamoyan, Hendrik Schuff, and Iryna Gurevych. 2024. LLM Roleplay: Simulating Human-Chatbot Interaction. arXiv:2407.03974 [cs.CL] https://arxiv. org/abs/2407.03974

    2024 arXiv

  23. [31]

    arXiv:2405.13003 [cs.CL] https://arxiv.org/abs/2405.13003

    A Survey on Recent Advances in Conversational Data Generation. arXiv:2405.13003 [cs.CL] https://arxiv.org/abs/2405.13003

    arXiv

  24. [32]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan...

    2024 arXiv

  25. [33]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy Lil...

    2025 arXiv

  26. [34]

    Canwen Xu, Daya Guo, Nan Duan, and Julian McAuley. 2023. Baize: An Open- Source Chat Model with Parameter-Efficient Tuning on Self-Chat Data. InProceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing , Houda Bouamor, Juan Pino, and Kalika Bali ...

    2023 doi

  27. [35]

    Bo-Hsiang Tseng, Yinpei Dai, Florian Kreyssig, and Bill Byrne. 2021. Transfer- able Dialogue Systems and User Simulators. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Interna- tional Joint Conference on Natural Languag...

    2021 doi

  28. [36]

    Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing Dialogue Agents: I have a dog, do you have pets too?. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) ,...

    2018 doi

  29. [37]

    Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. 2025. WizardLM: Em- powering large pre-trained language models to follow complex instructions. arXiv:2304.12244 [cs.CL] https://arxiv.org/abs/2304.12244

    2025 arXiv

  30. [38]

    Chujie Zheng, Sahand Sabour, Jiaxin Wen, Zheng Zhang, and Minlie Huang. 2023. AugESC: Dialogue Augmentation with Large Language Models for Emotional Support Conversation. arXiv:2202.13047 [cs.CL] https://arxiv.org/abs/2202.13047

    2023 arXiv

  31. [39]

    Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, and Yun- tian Deng. 2024. WildChat: 1M ChatGPT Interaction Logs in the Wild. arXiv:2405.01470 [cs.CL] https://arxiv.org/abs/2405.01470

    2024 arXiv

  32. [40]

    P Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT- Bench and Chatbot Arena. arXiv:2306.05685 [cs.CL]

    2023 arXiv

  33. [41]

    Xing, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle Li, Siyuan Zhuang, Zhang- hao Wu, Yonghao Zhuang, Zhuohan Li, Zi Lin, Eric P. Xing, Joseph E. Gonzalez, Ion Stoica, and Hao Zhang. 2024. LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset. arXiv:2309.11998 [cs.CL]...

    2024 arXiv

  34. [2024]

    arXiv:2410.21276 [cs.CL] https://arxiv.org/abs/2410

    GPT-4o System Card. arXiv:2410.21276 [cs.CL] https://arxiv.org/abs/2410. 21276

    arXiv

Pith tools

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