# WildChat: annotated research list Selected papers introducing, exploring, studying, or extending WildChat. Research hub: https://yuntiandeng.com/wildchat/ Research list: https://yuntiandeng.com/wildchat/research/ Release guide: https://yuntiandeng.com/wildchat/releases/ Dates below identify the first arXiv version unless the date note states otherwise. They do not establish the first public appearance of a project or dataset. The cited version and citation year are specified separately. ## WildChat: 1M ChatGPT Interaction Logs in the Wild Wenting Zhao; Xiang Ren; Jack Hessel; Claire Cardie; Yejin Choi; Yuntian Deng Wenting Zhao and Yuntian Deng contributed equally. - Venue: ICLR 2024 - Preprint date: 2024-05-02 - Date note: First arXiv version; this is not the date of WildChat's earlier ICLR submission or initial dataset release. - Date source: https://arxiv.org/abs/2405.01470 - Relationship: Original dataset - Topic: dataset - Data used: The original paper's approximately one-million-conversation release; some experiments use earlier collection windows. - Research record: https://yuntiandeng.com/wildchat/papers/wildchat/ - Paper: https://proceedings.iclr.cc/paper_files/paper/2024/hash/9421261e06f1a63a352b068f1ac90609-Abstract-Conference.html - arXiv: https://arxiv.org/abs/2405.01470 - Cited version: Cites the ICLR 2024 proceedings paper; the date above identifies its first arXiv version, not the beginning of the project. - Citation title: WildChat: 1M ChatGPT Interaction Logs in the Wild - Citation year: 2024 - Citation source: https://proceedings.iclr.cc/paper_files/paper/2024/hash/9421261e06f1a63a352b068f1ac90609-Abstract-Conference.html - Proceedings: International Conference on Learning Representations - BibTeX: https://yuntiandeng.com/wildchat/papers/wildchat/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/wildchat/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/wildchat/record.json Introduces consented, multilingual, multi-turn conversations with GPT-3.5 and GPT-4, together with timestamps and metadata for studying use across time and locations. The paper examines language and prompt diversity, toxicity and jailbreaks, and instruction tuning with WildLlama. Finding: The paper reports 68 languages with more than 100 user prompts each and demonstrates instruction tuning with a Llama-2 7B model. Evidence: https://arxiv.org/html/2405.01470v1 Sections 2–5 document collection, the language threshold, safety analyses, and the WildLlama experiment; Section 6 discusses sampling limitations. ## WildVis: Open Source Visualizer for Million-Scale Chat Logs in the Wild Yuntian Deng; Wenting Zhao; Jack Hessel; Xiang Ren; Claire Cardie; Yejin Choi - Venue: EMNLP 2024 System Demonstrations - Preprint date: 2024-09-05 - Date note: First arXiv version. - Date source: https://arxiv.org/abs/2409.03753 - Relationship: Exploration tool - Topic: tool - Data used: WildChat at approximately one million conversations, alongside LMSYS-Chat-1M; an exact WildChat revision is not specified. - Research record: https://yuntiandeng.com/wildchat/papers/wildvis/ - Paper: https://aclanthology.org/2024.emnlp-demo.50/ - arXiv: https://arxiv.org/abs/2409.03753 - Code and resources: https://github.com/da03/WildVisualizer - Cited version: Cites the EMNLP 2024 System Demonstrations proceedings paper. - Citation title: WildVis: Open Source Visualizer for Million-Scale Chat Logs in the Wild - Citation year: 2024 - Citation source: https://aclanthology.org/2024.emnlp-demo.50/ - Proceedings: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: System Demonstrations - DOI: https://doi.org/10.18653/v1/2024.emnlp-demo.50 - BibTeX: https://yuntiandeng.com/wildchat/papers/wildvis/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/wildvis/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/wildvis/record.json Provides composable search filters and an embedding map for inspecting WildChat and LMSYS-Chat-1M. Its case studies connect individual conversations to broader topic patterns and reproduce examples from research on journalism and personal disclosures. Finding: The visual comparison identifies a Midjourney prompt-generation cluster in WildChat and a chemistry cluster in LMSYS-Chat-1M. Evidence: https://arxiv.org/html/2409.03753v1 Sections 2–4 describe the interface, indexed datasets, and case studies; Appendix E expands the cross-dataset comparison. ## WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild Bill Yuchen Lin; Yuntian Deng; Khyathi Chandu; Faeze Brahman; Abhilasha Ravichander; Valentina Pyatkin; Nouha Dziri; Ronan Le Bras; Yejin Choi - Venue: ICLR 2025 - Preprint date: 2024-06-07 - Date note: First arXiv version; the paper also describes benchmark releases in March and May 2024. - Date source: https://arxiv.org/abs/2406.04770 - Relationship: Uses WildChat - Topic: evaluation - Data used: WildBench V2: 1,024 tasks curated from WildChat collection data and held out of the public WildChat training corpus. - Research record: https://yuntiandeng.com/wildchat/papers/wildbench/ - Paper: https://openreview.net/forum?id=MKEHCx25xp - arXiv: https://arxiv.org/abs/2406.04770 - Code and resources: https://github.com/allenai/WildBench - Cited version: Cites the arXiv preprint with its original 2024 preprint year. The paper later appeared at ICLR 2025. - Citation title: WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild - Citation year: 2024 - Citation source: https://arxiv.org/abs/2406.04770 - DOI: https://doi.org/10.48550/arXiv.2406.04770 - BibTeX: https://yuntiandeng.com/wildchat/papers/wildbench/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/wildbench/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/wildbench/record.json Turns challenging user requests from the WildChat collection into an evaluation set with task-specific checklists. The selection process combines language and quality filters, model difficulty estimates, and manual review, while reserving selected conversations from the public WildChat release. Finding: For the top-ranked models evaluated in the paper, WB-Reward has a Pearson correlation of 0.98 with Chatbot Arena's human-voted Elo ratings. Evidence: https://arxiv.org/html/2406.04770v2 Section 2.1 specifies selection and held-out data; Sections 3–4 define the metrics and compare model rankings. ## Trust No Bot: Discovering Personal Disclosures in Human-LLM Conversations in the Wild Niloofar Mireshghallah; Maria Antoniak; Yash More; Yejin Choi; Golnoosh Farnadi - Venue: COLM 2024 - Preprint date: 2024-07-16 - Date note: First arXiv version. - Date source: https://arxiv.org/abs/2407.11438 - Relationship: Uses WildChat - Topic: privacy - Data used: A 5,000-conversation English sample from the one-million-conversation WildChat corpus, limited to one conversation per hashed IP with additional prefix and length filters. - Research record: https://yuntiandeng.com/wildchat/papers/trust-no-bot/ - Paper: https://openreview.net/forum?id=tIpWtMYkzU - arXiv: https://arxiv.org/abs/2407.11438 - Code and resources: https://github.com/mireshghallah/ChatGPT-personal-disclosures - Cited version: Cites the arXiv preprint with its original 2024 preprint year. The paper also appeared at COLM 2024. - Citation title: Trust No Bot: Discovering Personal Disclosures in Human-LLM Conversations in the Wild - Citation year: 2024 - Citation source: https://arxiv.org/abs/2407.11438 - DOI: https://doi.org/10.48550/arXiv.2407.11438 - BibTeX: https://yuntiandeng.com/wildchat/papers/trust-no-bot/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/trust-no-bot/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/trust-no-bot/record.json Studies the tasks and contexts in which people disclose personal or sensitive information to chatbots. It combines automated PII detection with task and sensitive-topic taxonomies, releasing annotations that support analyses beyond conventional identifiers. Finding: In the analyzed sample, sensitive disclosures extend beyond the categories recognized by PII detectors, including preferences, personal habits, and application-related information. Evidence: https://arxiv.org/html/2407.11438v2 Section 2 specifies sampling; Sections 3–4 compare detected identifiers with sensitive topics. The observations apply to this filtered sample. ## Developing Story: Case Studies of Generative AI's Use in Journalism Natalie Grace Brigham; Chongjiu Gao; Tadayoshi Kohno; Franziska Roesner; Niloofar Mireshghallah - Venue: SoLaR Workshop at NeurIPS 2024 - Preprint date: 2024-06-19 - Date note: First arXiv version, initially titled Breaking News; the revised title is Developing Story. - Date source: https://arxiv.org/abs/2406.13706 - Relationship: Uses WildChat - Topic: behavior - Data used: An earlier WildChat release containing approximately 650,000 conversations, before the one-million-conversation release. - Research record: https://yuntiandeng.com/wildchat/papers/developing-story/ - Paper: https://openreview.net/forum?id=415AZy8Jih - arXiv: https://arxiv.org/abs/2406.13706 - Cited version: Cites the arXiv preprint with its original 2024 preprint year. Uses the revised Developing Story title; the paper also appeared at SoLaR 2024. - Citation title: Developing Story: Case Studies of Generative AI's Use in Journalism - Citation year: 2024 - Citation source: https://arxiv.org/abs/2406.13706 - DOI: https://doi.org/10.48550/arXiv.2406.13706 - BibTeX: https://yuntiandeng.com/wildchat/papers/developing-story/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/developing-story/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/developing-story/record.json Matches conversations in WildChat to published articles from two news agencies, documenting how source material and generated drafts enter journalistic workflows. The study examines input materials, overlap with published text, and publication timing; its cases do not establish prevalence across journalism. Finding: The authors match 78 conversations to 111 published articles and report a median output-to-publication ROUGE-L overlap of 0.62. Evidence: https://arxiv.org/html/2406.13706v2 Section 2 identifies the earlier release; Table 1 reports verified matches; Section 3 reports text overlap. Section 5 states that identified sensitive conversations were subsequently removed. ## WildChat-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training Benjamin Feuer; Chinmay Hegde - Venue: ICML 2025 - Preprint date: 2025-01-30 - Date note: First arXiv version. - Date source: https://arxiv.org/abs/2501.18511 - Relationship: Extends WildChat - Topic: training - Data used: A third-party synthetic-response extension of WildChat-1M, retaining its user prompts and adding responses from open-weight models. - Research record: https://yuntiandeng.com/wildchat/papers/wildchat-50m/ - Paper: https://proceedings.mlr.press/v267/feuer25a.html - arXiv: https://arxiv.org/abs/2501.18511 - Code and resources: https://github.com/penfever/wildchat-50m - Cited version: Cites the ICML 2025 proceedings paper in PMLR volume 267. - Citation title: WildChat-50M: A Deep Dive Into the Role of Synthetic Data in Post-Training - Citation year: 2025 - Citation source: https://proceedings.mlr.press/v267/feuer25a.html - Proceedings: Proceedings of the 42nd International Conference on Machine Learning - BibTeX: https://yuntiandeng.com/wildchat/papers/wildchat-50m/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/wildchat-50m/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/wildchat-50m/record.json Uses WildChat prompts to compare responses from more than 50 open-weight model checkpoints and study the choice of synthetic-data generator. The authors also construct Re-Wild, a supervised fine-tuning mixture, and test it with Llama-3.1 8B. Finding: The reported Re-Wild mixture outperforms the Tulu-3 SFT mixture on the paper's aggregate evaluations using 40% as many training samples. Evidence: https://arxiv.org/html/2501.18511v1 Section 2 identifies WildChat-1M as the prompt source and describes 54 checkpoints; Section 3 reports the SFT mixtures and evaluations. ## WildClaims: Conversational Information Access in the Wild(Chat) Hideaki Joko; Shakiba Amirshahi; Charles L. A. Clarke; Faegheh Hasibi - Venue: ECIR 2026 - Preprint date: 2025-09-22 - Date note: First arXiv version; the ECIR 2026 paper uses the revised title shown here. - Date source: https://arxiv.org/abs/2509.17442 - Relationship: Uses WildChat - Topic: information - Data used: 3,000 sampled English WildChat conversations after excluding math and coding; the paper does not identify an exact dataset revision. - Research record: https://yuntiandeng.com/wildchat/papers/wildclaims/ - Paper: https://link.springer.com/chapter/10.1007/978-3-032-21321-1_47 - arXiv: https://arxiv.org/abs/2509.17442 - Code and resources: https://github.com/shakibaam/wildclaims - Cited version: Cites the published ECIR 2026 paper under its revised title; the preprint first appeared in September 2025. - Citation title: WildClaims: Conversational Information Access in the Wild(Chat) - Citation year: 2026 - Citation source: https://link.springer.com/chapter/10.1007/978-3-032-21321-1_47 - Proceedings: Advances in Information Retrieval. ECIR 2026 - DOI: https://doi.org/10.1007/978-3-032-21321-1_47 - BibTeX: https://yuntiandeng.com/wildchat/papers/wildclaims/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/wildclaims/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/wildclaims/record.json Examines factual information conveyed by chatbot responses, including when the user's main task is editing or creative writing. It releases extracted claims and check-worthiness annotations to study information access beyond explicit questions. Finding: WildClaims contains 121,905 extracted claims from 7,587 assistant utterances; estimated check-worthy-conversation prevalence varies substantially with the annotation method. Evidence: https://arxiv.org/html/2509.17442v2 Section 3 and Table 2 specify selection and annotation counts; the prevalence analysis distinguishes conservative and more permissive classification choices. ## Adopt ≠ Adapt: Longitudinal Analyses of LLM Conversations in the Wild Rebecca M. M. Hicke; Kiran Tomlinson - Venue: arXiv preprint - Preprint date: 2026-05-27 - Date note: First arXiv version; annotation follows the August 2026 revision. - Date source: https://arxiv.org/abs/2605.29018 - Relationship: Uses WildChat - Topic: behavior - Data used: WildChat-4.8M, with hashed-IP filters and a main-analysis cutoff before September 2024 to reduce later API-like activity. - Research record: https://yuntiandeng.com/wildchat/papers/adopt-adapt/ - Paper: https://arxiv.org/abs/2605.29018 - arXiv: https://arxiv.org/abs/2605.29018 - Code and resources: https://github.com/microsoft/adopt-adapt-wildchat - Cited version: Cites the arXiv preprint with its original 2026 preprint year. The annotation follows the August 2026 revision. - Citation title: Adopt ≠ Adapt: Longitudinal Analyses of LLM Conversations in the Wild - Citation year: 2026 - Citation source: https://arxiv.org/abs/2605.29018 - DOI: https://doi.org/10.48550/arXiv.2605.29018 - BibTeX: https://yuntiandeng.com/wildchat/papers/adopt-adapt/citation.bib - CSL JSON: https://yuntiandeng.com/wildchat/papers/adopt-adapt/citation.csl.json - Record JSON: https://yuntiandeng.com/wildchat/papers/adopt-adapt/record.json Compares individual and population-level usage patterns in Bing Copilot with analyses of WildChat-4.8M. The study separates changes over a user's trajectory from differences between users, and uses the cross-platform comparison to examine how WildChat's sample differs. Finding: WildChat users' intent and domain distributions more closely resemble high-activity Copilot users than low-activity users; hashed IPs remain an imperfect proxy for individuals. Evidence: https://arxiv.org/html/2605.29018v2 Section 3.2 describes the WildChat filters and time cutoff; Section 6 and Figure 6 compare activity groups across the two sources. ## Complete exports - Records: https://yuntiandeng.com/wildchat/data/papers.json - BibTeX: https://yuntiandeng.com/wildchat/data/papers.bib - CSL JSON: https://yuntiandeng.com/wildchat/data/papers.csl.json - Releases: https://yuntiandeng.com/wildchat/data/releases.json - Navigation: https://yuntiandeng.com/wildchat/data/navigation.json