Pith. sign in

REVIEW 4 major objections 4 minor 28 references

ProfiLLM: An LLM-Based Framework for Implicit Profiling of Chatbot Users

T0 review · 4 major / 4 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 ProfiLLM claims that an LLM can infer a chatbot user's technical proficiency in 23 IT/cybersecurity subdomains solely from ordinary troubleshooting prompts, cutting the predicted-vs-actual gap by 55–65% after a single prompt.

desk verdict Useful package, but the headline result is per-subdomain after an on-topic prompt, not after one user turn, and the synthetic evaluation loop leaves generalization unproven. read the letter →

arxiv 2506.13980 v1 pith:EXO2YP63 submitted 2025-06-16 cs.AI

Shahaf David , Yair Meidan , Ido Hersko , Daniel Varnovitzky , Dudu Mimran , Yuval Elovici , Asaf Shabtai This is my paper
classification cs.AI
keywords userprofilingLLMchatbotsimplicitITSectaxonomypersonasimulationproficiencyinference
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

ProfiLLM is a framework for implicit, dynamic user profiling: it infers a chatbot user's expertise from the natural-language prompts they type, with no questionnaires and no self-reports. Applied to IT and cybersecurity, the paper defines 23 proficiency subdomains grouped into five domains and reports that after a single user prompt the gap between predicted and actual proficiency scores shrinks by 55–65 percent, with mean absolute error converging to the 0.3–0.65 range on a 1–5 scale. If this holds, a chatbot could adjust jargon, explanation depth, and step-by-step detail to each user's evolving skill level during an ongoing conversation. The primary evidence is 1,315 high-quality conversations from 263 synthetic users seeded by 63 human questionnaire responses, complemented by a five-person human pilot.

What carries the argument

The central machinery is a taxonomy plus a confidence-weighted update rule. The taxonomy consists of 23 ITSec subdomains across five domains (Hardware, Networking, Cybersecurity, Software, Operating Systems); each prompt is assigned to one or more subdomains, so the profiler always knows which expertise dimension it is scoring. For each assigned subdomain, the LLM produces a temporary score $p^{\mathrm{temp},i}_j$, and the running profile is updated as $p^i_j = \alpha^i_j p^{\mathrm{temp},i}_j + (1-\alpha^i_j)p^{i-1}_j$, with $\alpha^i_j = \alpha_0/(1+\beta i)$. The inverse time decay makes early prompts count heavily while the estimate is unreliable and later prompts count less as confidence grows; the paper also evaluated fixed weights and concurrent scoring and found the decaying, per-subdomain version best.

What would settle it

Have at least 100 real users from varied backgrounds complete the ProfiLLM questionnaire, then troubleshoot with the ITSec chatbot, and compute per-subdomain MAE after the first prompt assigned to each subdomain. If the error relative to the default score of 3 does not drop by roughly 55–65% (or if later prompts do not converge toward the 0.3–0.65 range), the synthetic evaluation does not transfer to real users.

Watch

Extended reading notes

Core claim

ProfiLLM[ITSec] claims that a general-purpose LLM can serve as an online proficiency profiler: it reads each troubleshooting prompt, assigns it to relevant taxonomy subdomains, scores the prompt's implied knowledge on a 1–5 scale using concept complexity, terminology, depth of understanding, and conversational context, and folds the new score into a running per-subdomain estimate. With the optimized configuration—GPT-4o as the profiler, a context window of one previous exchange, and decay rate β=0.1—the mean absolute error falls from about 1.14 at iteration 0 to roughly 0.5 after a single prompt per subdomain, the reported 55–65% gap reduction for novice and advanced archetypes, then fluctuates and settles in the 0.3–0.65 range for domains. Intermediate-archetype users show an early error increase before stabilizing, which the paper attributes to their less distinctive language. The core discovery is that rapid and accurate profiling can be fully implicit: no self-assessments, only the user's ordinary troubleshooting language.

Load-bearing premise

The load-bearing premise is that LLM-generated synthetic users and their conversations behave like real human users; if that premise is false, the reported 55–65% one-prompt improvement may not appear with real chatbot users.

Editorial extensions

If this is right

  • A help-desk or support chatbot could adapt vocabulary and step granularity per user without ever asking for a self-assessment, since the profiler works on ordinary prompts.
  • The framework is domain-adaptable: creating a new profiler reduces to defining a taxonomy and tuning system prompts, so legal, medical, or data-science expertise could be profiled the same way.
  • Because most of the accuracy gain comes after one prompt, response adaptation could start almost immediately in a live conversation, with further refinement as more prompts arrive.
  • Continued interaction converges per-subdomain MAE to the 0.3–0.65 range, which the paper argues is practically negligible for choosing appropriate response complexity.
  • The synthetic persona pipeline provides a reusable method for building labeled conversational profiling datasets in domains where real interaction logs are scarce.

Reading between the lines

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

  • The evaluation mixes the same model family into persona generation, conversation simulation, and profiling, so the one-prompt improvement may partly reflect the profiler recognizing its own synthetic stereotypes; the five-person human pilot is too small to rule this out, and real-user scale-up is the unresolved test.
  • Because scores update only for subdomains a user actually touches, unmentioned subdomains keep the neutral default of 3; a user who never discusses malware is silently treated as average there, making the implicit profile uneven even when per-subdomain MAE looks good.
  • The decay rule treats all prompts as equally informative after the same count; a natural untested extension is to weight prompts by informativeness or by disagreement with the current estimate, which could reduce the early fluctuations the paper observes.
  • A concrete transfer test would be to use different model families for persona generation versus profiling; if the 55–65% one-prompt gain collapses under that swap, the result is likely synthetic-data leakage rather than a general property of implicit conversational profiling.
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

4 major / 4 minor

Summary. The paper proposes ProfiLLM, an LLM-based framework for implicit, dynamic user profiling in chatbot conversations. The framework combines a domain-adapted taxonomy with an LLM module that assigns each user prompt to subdomains, scores the prompt per assigned subdomain, and updates profile scores via a decaying weighted average (Eqs. 1 and 2). The authors instantiate the framework for IT/cybersecurity (ITSec) with a 23-subdomain taxonomy, generate synthetic users from cluster centroids of 63 human questionnaire responses, use GPT-4o to create personas and chatbot conversations, filter conversations with Claude, and evaluate on 1,315 conversations from 263 synthetic users. They report rapid MAE reductions, up to 55–65% after one prompt for novice and advanced archetypes, plus a five-participant human pilot. The paper also contributes the taxonomy, a persona simulation methodology, an ablation study, and a dataset/code release.

Significance. If the reported accuracy held for real users, the framework would be a useful contribution: it offers fully implicit, dynamic profiling with no questionnaire burden, a reusable domain-adapted taxonomy, and an LLM-based persona simulation methodology. The paper is transparent about the per-subdomain definition of MAE@i, and it includes an ablation, an archetype analysis, a prompt-length analysis, and a small human experiment, which are all valuable evaluation components. However, the central quantitative claims rest on a synthetic evaluation loop in which the same model family generates the personas and conversations that are later profiled, and the hyperparameters are tuned on users who appear to be included in the final evaluation. These issues mean the headline numbers are not yet established for real-world deployment; the contribution is promising but the evidence is currently insufficient to support the abstract's claims as stated.

major comments (4)
  1. [Section 4.2 / Section 4.3] The hyperparameter selection appears to contaminate the evaluation. Section 4.2 performs a grid search over LLM, beta, and |W| on a sample of 30 randomly selected users, and Section 4.3 then applies the optimized configuration to the entire refined experimental dataset. The paper does not state that the 30 tuning users are excluded from the 263 users in the final evaluation, so the MAE values in Sections 4.3 and 4.4 may be optimistically biased. Please use a proper held-out split (for example, tune on the 30 users but report evaluation on the remaining 233), or report nested cross-validation results.
  2. [Sections 3.3, 3.4, and 4.1] The evaluation's external validity is a major concern. The synthetic ground-truth profiles are derived from cluster centroids of 63 human questionnaires (Section 3.2), but the persona descriptions are generated by GPT-4o (Section 3.3), the conversations are generated by GPT-4o (Section 3.4), and the profiler is also GPT-4o (Section 3.1). High accuracy on this pipeline may partly reflect the model's ability to recognize its own stereotypes rather than a genuine ability to infer proficiency from human troubleshooting language. The five-participant human experiment in Section 4.7 is too small to rule this out. Please add an evaluation on independently collected human conversations, or at a minimum use a different model family for data generation versus profiling and report the difference.
  3. [Abstract / Sections 4.2 and 4.4] The abstract's claim of reducing the gap by “up to 55–65% after a single prompt” is not supported at the level of a single user message. Section 4.2 defines MAE@i as the error after the i-th prompt assigned within each subdomain, and Section 4.4 reports the 55–65% reduction at iteration 1 only for novice and advanced users and only per subdomain; Section 4.3 reports an average reduction of about 30%, and intermediate users' MAE increases by 75% at iteration 1. Since a real user turn is typically assigned to only one or a few of the 23 subdomains, the whole-profile improvement after one prompt is much smaller than the advertised range. Please rephrase the abstract and conclusions to state the per-subdomain, archetype-selective nature of the result.
  4. [Abstract / Section 4.1] The abstract says the method was “evaluated on 1,760 human-like chatbot conversations from 263 synthetic users,” but Section 4.1 reports that 352 synthetic users generated 1,760 conversations, and after applying an 8.5 quality threshold the final dataset comprised 1,315 conversations from 263 users. Please align the abstract with the final filtered dataset, or clearly report both pre- and post-filter numbers in the abstract.
minor comments (4)
  1. [Section 3.4] The sentence “another LLM valuated the conversations” contains a typo; it should read “evaluated.”
  2. [Section 4.5] The statement that there is “no statistically significant correlation” would benefit from the actual correlation coefficients and p-values, or a confidence interval, rather than only the r values shown in Figure 4b.
  3. [Section 3.3 / Footnote 1] The paper repeatedly mentions publicly releasing the codebase and dataset, but Footnote 1 says the repository URL is “to be shared upon request.” Please either provide the URL or qualify the availability statement.
  4. [Section 4.2 / Table 2] The grid search fixes alpha_0 = 0.8, but the paper does not discuss whether alpha_0 was also optimized or how sensitive the results are to this choice; a sentence explaining this would improve completeness.

Circularity Check

1 steps flagged · score 6.0 of 10

Central 55–65% result rests on a same-model synthetic loop: GPT-4o converts profile vectors into personas, GPT-4o generates the conversations, and GPT-4o (via ProfiLLM) infers the profiles.

  1. fitted input called prediction [Sec. 3.3–3.4, Sec. 4.1, Sec. 4.4 (synthetic data generation, quality filtering, and archetype performance)]
    "we developed a prompt that converts a numeric profile vector into a textual persona description. ... The (synthetic) users were instructed to maintain consistent behavior and phrase their prompts strictly in accordance with their assigned persona. ... another LLM valuated the conversations based on two criteria: the alignment between the user’s profile and their behavior/phrasing and the naturalness of the conversation flow. ... Conversations not meeting either criterion were excluded from the dataset."

    The 'actual' ITSec profile for each synthetic user is the numeric vector that GPT-4o is prompted to convert into a persona; the same GPT-4o backend then generates the user's troubleshooting prompts and is also the model behind ProfiLLM's scoring. Thus the conversation text is produced by conditioning GPT-4o on the very label that GPT-4o is later asked to predict, so the reported MAE reductions largely measure the model's ability to decode its own stereotype-encoding. The quality filter further removes conversations where behavior does not align with the profile, selecting for cases where the label is most recoverable. The only independent check (Sec. 4.7, five graduate students, one scenario) is too small to support the headline 55–65% improvement on real users.

full rationale

The paper's headline quantitative claim is not circular in an equation-level sense, but its main evaluation loop is self-referential: GPT-4o generates the persona from the ground-truth profile vector, GPT-4o generates the conversation from that persona, and GPT-4o (through ProfiLLM) infers the profile from the conversation. The paper does report a small human-user experiment as an external anchor, and the MAE metric is defined transparently; however, five participants and one scenario do not break the same-model loop. No load-bearing self-citation or imported uniqueness theorem is present. The circularity is partial but central: the reported 55–65% 'after a single prompt' improvement is demonstrated on synthetic data whose labels and language both originate from the model being evaluated.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends mainly on the untested transferability of LLM-generated synthetic conversations to real users, plus a small number of tuned hyperparameters. No new physical or conceptual entities are introduced beyond the taxonomy and persona archetypes, which are data-derived rather than postulated.

free parameters (5)
  • beta (decay rate) = 0.1
    Optimized by grid search over {0.1, 0.2, 0.3} on 30 random users (Table 2).
  • |W| (context window size) = 1
    Optimized by grid search over {0, 1, 999} on the same 30 users.
  • alpha0 (initial confidence weight) = 0.8
    Chosen, not optimized; used in Eq. 2 for all experiments.
  • Conversation quality threshold = 8.5
    Used to filter synthetic conversations (Sec. 4.1); value chosen without sensitivity analysis.
  • Number of clusters k = 3
    Selected by Elbow and Silhouette methods (Sec. 4.1), giving novice/intermediate/advanced archetypes.
assumptions (5)
  • domain assumption GPT-4o produces reliable and consistent profile scores on a 5-point scale across different calls and contexts.
    The entire profile inference rests on the LLM's scoring of concept complexity, terminology, depth, and context (Sec. 2.2).
  • domain assumption Self-reported questionnaire scores are a valid ground truth for technical proficiency.
    Human profiles used to create clusters and synthetic users come from a 23-subdomain questionnaire relying on self-assessment and self-report (Sec. 3.2).
  • domain assumption LLM-generated personas and conversations preserve the numeric proficiency profile in observable language.
    Synthetic users are generated from persona archetypes and their conversations are filtered by an LLM quality check; this assumes the profile is reflected in wording (Sec. 3.3, 3.4).
  • domain assumption The 23-subdomain taxonomy partitions ITSec troubleshooting dialogue with no major omissions.
    Taxonomy was built from syllabi, forums, and surveys (Sec. 2.1); any unlisted subdomain would not be scored.
  • standard math Weighted averaging with inverse time decay converges to a stable profile under repeated prompts.
    Eq. 1-2 define the update; the convergence behavior is observed empirically, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProfiLLM: An LLM-Based Framework for Implicit Profiling of Chatbot Users." pith.science (2026). https://pith.science/paper/EXO2YP63

@misc{pith2026250613980,
  author       = {Pith},
  title        = {Pith review of: ProfiLLM: An LLM-Based Framework for Implicit Profiling of Chatbot Users},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EXO2YP63}},
  note         = {Machine review of arXiv:2506.13980}
}
read the original abstract

Despite significant advancements in conversational AI, large language model (LLM)-powered chatbots often struggle with personalizing their responses according to individual user characteristics, such as technical expertise, learning style, and communication preferences. This lack of personalization is particularly problematic in specialized knowledge-intense domains like IT/cybersecurity (ITSec), where user knowledge levels vary widely. Existing approaches for chatbot personalization primarily rely on static user categories or explicit self-reported information, limiting their adaptability to an evolving perception of the user's proficiency, obtained in the course of ongoing interactions. In this paper, we propose ProfiLLM, a novel framework for implicit and dynamic user profiling through chatbot interactions. This framework consists of a taxonomy that can be adapted for use in diverse domains and an LLM-based method for user profiling in terms of the taxonomy. To demonstrate ProfiLLM's effectiveness, we apply it in the ITSec domain where troubleshooting interactions are used to infer chatbot users' technical proficiency. Specifically, we developed ProfiLLM[ITSec], an ITSec-adapted variant of ProfiLLM, and evaluated its performance on 1,760 human-like chatbot conversations from 263 synthetic users. Results show that ProfiLLM[ITSec] rapidly and accurately infers ITSec profiles, reducing the gap between actual and predicted scores by up to 55--65\% after a single prompt, followed by minor fluctuations and further refinement. In addition to evaluating our new implicit and dynamic profiling framework, we also propose an LLM-based persona simulation methodology, a structured taxonomy for ITSec proficiency, our codebase, and a dataset of chatbot interactions to support future research.

Figures

Figures reproduced from arXiv: 2506.13980 by the authors.

Figure 1
Figure 1. User Profile Inference in Terms of the Taxonomy [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Experimental data collected: cluster centroids and conversation quality [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. P rof iLLMIT Sec ∗ ’s performance across profile domains and subdomains [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Profiling performance by persona archetype and prompt length [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Profiling performance in the ablation study and human users experiment [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 22 canonical work pages

  1. [1]

    Let the llms talk: Simulating human-to-human conversational qa via zero-shot llm-to-llm interactions

    Zahra Abbasiantaeb, Yifei Yuan, Evangelos Kanoulas, and Mohammad Alianne- jadi. Let the llms talk: Simulating human-to-human conversational qa via zero-shot llm-to-llm interactions. InProceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 8–17, 2024

    work page 2024

  2. [2]

    The best it advice

    Electric AI. The best it advice. Electric AI Blog, [Online], 2025. Available: https://www.electric.ai/blog/the-best-it-advice

    work page 2025

  3. [3]

    Taxonomy of mobile users’ security awareness.Computers & Security, 73:266–293, 2018

    Ron Bitton, Andrey Finkelshtein, Lior Sidi, Rami Puzis, Lior Rokach, and Asaf Shabtai. Taxonomy of mobile users’ security awareness.Computers & Security, 73:266–293, 2018

    work page 2018

  4. [4]

    Diaggpt: An llm-based and multi-agent dialogue system with au- tomatic topic management for flexible task-oriented dialogue.arXiv preprint arXiv:2308.08043, 2023

    Lang Cao. Diaggpt: An llm-based and multi-agent dialogue system with au- tomatic topic management for flexible task-oriented dialogue.arXiv preprint arXiv:2308.08043, 2023

    arXiv 2023

  5. [5]

    Apollonion: Profile- centric dialog agent.arXiv preprint arXiv:2404.08692, 2024

    Shangyu Chen, Zibo Zhao, Yuanyuan Zhao, and Xiang Li. Apollonion: Profile- centric dialog agent.arXiv preprint arXiv:2404.08692, 2024

    arXiv 2024

  6. [6]

    Technical support fundamentals, dell technologies technical customer support specialist, ibm technical support, google it support, 2025

    Coursera. Technical support fundamentals, dell technologies technical customer support specialist, ibm technical support, google it support, 2025. Available: https://www.coursera.org

    work page 2025

  7. [7]

    The use of chatgpt for personality research: Administering questionnaires using generated personas

    Joost CF de Winter, Tom Driessen, and Dimitra Dodou. The use of chatgpt for personality research: Administering questionnaires using generated personas. Personality and Individual Differences, 228:112729, 2024

    work page 2024

  8. [8]

    Personality structure: Emergence of the five-factor model.Annual review of psychology, 41(1):417–440, 1990

    John M Digman. Personality structure: Emergence of the five-factor model.Annual review of psychology, 41(1):417–440, 1990

    work page 1990

Show all 28 references
  1. [9]

    User profiling and satisfaction inference in public information access services.Journal of Intelligent Information Systems, 58(1):67–89, 2022

    Arthur Marçal Flores, Matheus Camasmie Pavan, and Ivandré Paraboni. User profiling and satisfaction inference in public information access services.Journal of Intelligent Information Systems, 58(1):67–89, 2022

    2022

  2. [10]

    Martín, Alberto Fernández-Isabel, Isaac Martín de Diego, and Marta Beltrán

    Alejandro G. Martín, Alberto Fernández-Isabel, Isaac Martín de Diego, and Marta Beltrán. A survey for user behavior analysis based on machine learning techniques: current models and applications.Applied Intelligence, 51(8):6029–6055, 2021

    2021

  3. [11]

    Towards adaptive learning: Ai chatbot for users profiling and learning styles.International Journal on Technical and Physical Problems of Engineering, 16(59):167–174, 2024

    R Gandoul, N Chafiq, and M Ghazouani. Towards adaptive learning: Ai chatbot for users profiling and learning styles.International Journal on Technical and Physical Problems of Engineering, 16(59):167–174, 2024. 16 Shahaf David et al

    2024

  4. [12]

    Invest Implications: Forecast Analysis: Artificial Intelli- gence Software, 2023-2027, Worldwide, 2023

    Gartner. Invest Implications: Forecast Analysis: Artificial Intelli- gence Software, 2023-2027, Worldwide, 2023. [Online]. Available: https://www.gartner.com/en/documents/4925331. [Accessed: 28-Jan-2024]

    2023

  5. [13]

    Chatgpt for suicide risk assessment on social media: quantitative evaluation of model performance, potentials and limitations.arXiv preprint arXiv:2306.09390, 2023

    Hamideh Ghanadian, Isar Nejadgholi, and Hussein Al Osman. Chatgpt for suicide risk assessment on social media: quantitative evaluation of model performance, potentials and limitations.arXiv preprint arXiv:2306.09390, 2023

    2023 arXiv

  6. [14]

    Are llms all you need for task-oriented dia- logue?arXiv preprint arXiv:2304.06556, 2023

    Vojtěch Hudeček and Ondřej Dušek. Are llms all you need for task-oriented dia- logue?arXiv preprint arXiv:2304.06556, 2023

    2023 arXiv

  7. [15]

    Personality bert: a transformer- based model for personality detection from textual data

    Dipika Jain, Akshi Kumar, and Rohit Beniwal. Personality bert: a transformer- based model for personality detection from textual data. InProceedings of interna- tional conference on computing and communication networks: ICCCN 2021, pages 515–522. Springer, 2022

    2021

  8. [16]

    Cansu Koyuturk, Mona Yavari, Emily Theophilou, Sathya Bursic, Gregor Don- abauer, Alessia Telari, Alessia Testa, Raffaele Boiano, Alessandro Gabbiadini, Davinia Hernandez-Leo, et al. Developing effective educational chatbots with chat- gpt prompts: insights from preliminary te...

    2023 arXiv

  9. [17]

    Content-oriented user modeling for personalized response ranking in chatbots.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 26(1):122–133, 2017

    Bingquan Liu, Zhen Xu, Chengjie Sun, Baoxun Wang, Xiaolong Wang, Derek F Wong, and Min Zhang. Content-oriented user modeling for personalized response ranking in chatbots.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 26(1):122–133, 2017

    2017

  10. [18]

    Once:Boostingcontent- based recommendation with both open-and closed-source large language models

    QijiongLiu,NuoChen,TetsuyaSakai,andXiao-MingWu. Once:Boostingcontent- based recommendation with both open-and closed-source large language models. InProceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 452–461, 2024

    2024

  11. [19]

    One chatbot per person: Creating personalized chatbots based on implicit user profiles

    Zhengyi Ma, Zhicheng Dou, Yutao Zhu, Hanxun Zhong, and Ji-Rong Wen. One chatbot per person: Creating personalized chatbots based on implicit user profiles. InProceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages ...

    2021

  12. [20]

    Adaptive intelligence: Gpt-powered language models for dynamic re- sponses to emerging healthcare challenges.IJARCCE, 13:104–109, 01 2024

    Karthik Meduri, Hari Gonaygunta, Geeta Nadella, Priyanka Pawar, and Deepak Kumar. Adaptive intelligence: Gpt-powered language models for dynamic re- sponses to emerging healthcare challenges.IJARCCE, 13:104–109, 01 2024

    2024

  13. [21]

    Llm-based chatbots in language learning.European Jour- nal of Education, 7(1):102–123, 2024

    Panagiotis Panagiotidis. Llm-based chatbots in language learning.European Jour- nal of Education, 7(1):102–123, 2024

    2024

  14. [22]

    Learning implicit user profile for personalized retrieval-based chatbot

    Hongjin Qian, Zhicheng Dou, Yutao Zhu,Yueyuan Ma,and Ji-Rong Wen. Learning implicit user profile for personalized retrieval-based chatbot. Inproceedings of the 30th ACM international conference on Information & Knowledge Management, pages 1467–1477, 2021

    2021

  15. [23]

    Digcomp 2.2: The digital competence frame- work for citizens - with new examples of knowledge, skills and attitudes

    Vuorikari R, Kluzer S, and Punie Y. Digcomp 2.2: The digital competence frame- work for citizens - with new examples of knowledge, skills and attitudes. Scientific analysis or review, Joint Research Centre (JRC), European Commission, Luxem- bourg (Luxembourg), 2022

    2022

  16. [24]

    Leveraging gpt-4 capabilities for developing context- aware, personalized chatbot interfaces in e-commerce customer support systems

    Kathari Santosh, Timur Kholmukhamedov, M Sandeep Kumar, Mohd Aarif, Iskan- dar Muda, and B Kiran Bala. Leveraging gpt-4 capabilities for developing context- aware, personalized chatbot interfaces in e-commerce customer support systems. In2024 10th International Conference on C...

    2024

  17. [25]

    Llm roleplay: Simulat- ing human-chatbot interaction.arXiv preprint arXiv:2407.03974, 2024

    Hovhannes Tamoyan, Hendrik Schuff, and Iryna Gurevych. Llm roleplay: Simulat- ing human-chatbot interaction.arXiv preprint arXiv:2407.03974, 2024

    2024 arXiv

  18. [26]

    User modeling in the era of large language models: Current research and future directions.arXiv preprint arXiv:2312.11518, 2023

    Zhaoxuan Tan and Meng Jiang. User modeling in the era of large language models: Current research and future directions.arXiv preprint arXiv:2312.11518, 2023. P rof iLLM: LLM-based Framework for Implicit Profiling of Chatbot Users 17

    2023 arXiv

  19. [27]

    Desprompt: Personality-descriptive prompt tuning for few-shot personality recognition.Information Processing & Management, 60(5):103422, 2023

    Zhiyuan Wen, Jiannong Cao, Yu Yang, Haoli Wang, Ruosong Yang, and Shuaiqi Liu. Desprompt: Personality-descriptive prompt tuning for few-shot personality recognition.Information Processing & Management, 60(5):103422, 2023

    2023

  20. [28]

    When chatbots fail: exploring user coping following a chatbots-induced service failure.Information technology & people, 37(8):175–195, 2024

    Ruby Wenjiao Zhang, Xiaoning Liang, and Szu-Hsin Wu. When chatbots fail: exploring user coping following a chatbots-induced service failure.Information technology & people, 37(8):175–195, 2024

    2024

Pith tools

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