PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on LLM-based Conversational User Simulation

Bo Ni, Leyao Wang, Yu Wang, Branislav Kveton, Franck Dernoncourt, Yu Xia, Hongjie Chen, Reuben Leura, Samyadeep Basu, Subhojyoti Mukherjee, Puneet Mathur, Nesreen Ahmed, Junda Wu, Li Li, Huixin Zhang, Ruiyi Zhang, Tong Yu, Sungchul Kim, Jiuxiang Gu, Zhengzhong Tu, Alexa Siu, Zichao Wang, David Seunghyun Yoon, Nedim Lipka, Namyong Park, Zihao Lin, Trung Bui, Yue Zhao, Tyler Derr, Ryan A. Rossi

arXiv:2604.24977v1cs.CLcs.HC

TL;DR

Conversational user simulation must capture the interactive nature of human communication, yet existing research is dispersed across methods and evaluation practices. This survey organizes the field through a unified taxonomy, synthesizes techniques and protocols, and identifies open challenges for future research.

  • Problem

    Existing user-simulation methods do not fully capture the interactive, conversational nature and complexity of human communication.

  • Method

    The survey organizes LLM-based conversational user simulation through a unified Who, What, and How framework, covering methods, applications, evaluations, and datasets.

  • Results

    The survey categorizes representative literature, examines evaluation protocols and datasets, and identifies open challenges for more consistent, diverse, and trustworthy user simulators.

  • Takeaways & Limitations

    The unified framework supports benchmarking and future research on LLM-based conversational user simulation.

  • Takeaways & Limitations

    The survey does not provide a full benchmarking study across methods, and some hybrid or domain-specific methods may not fit its taxonomy.

Abstract

from arXiv · show

User simulation has long played a vital role in computer science due to its potential to support a wide range of applications. Language, as the primary medium of human communication, forms the foundation of social interaction and behavior. Consequently, simulating conversational behavior has become a key area of study. Recent advancements in large language models (LLMs) have significantly catalyzed progress in this domain by enabling high-fidelity generation of synthetic user conversation. In this paper, we survey recent advancements in LLM-based conversational user simulation. We introduce a novel taxonomy covering user granularity and simulation objectives. Additionally, we systematically analyze core techniques and evaluation methodologies. We aim to keep the research community informed of the latest advancements in conversational user simulation and to further facilitate future research by identifying open challenges and organizing existing work under a unified framework.

1 Introduction

LLMs have transformed user simulation by enabling general-purpose, contextually rich simulated interactions with less data and task-specific training, motivating focused study of conversational user simulation. This survey provides a unified taxonomy organized around who, what, and how users and conversations are simulated, while identifying research trends and open challenges.

  • Background: Prior user simulation spans statistical preference and behavior models, simulation games, agent-based LLM environments, retrieval-augmented generation, and chain-of-thought reasoning.These approaches emulate preferences, behaviors, knowledge review and internalization, or step-by-step human-like deliberation.
  • Motivation: LLMs enable general-purpose simulation across tasks and domains while lowering the barrier to generating high-quality, contextually rich interactions through prompt engineering.Earlier recommender-system simulators required large user datasets and were tailored to particular forms of simulation.
  • Motivation: Realistic user interactions require explicit modeling of communication’s interactive and conversational nature, creating an opportunity for LLM-based conversational user simulation.Language is described as the primary medium of human interaction, and LLMs provide unprecedented language capability.
  • Survey scope: The survey offers a comprehensive overview and unified taxonomy of conversational user simulation, defining its key components and scope.The taxonomy is presented as an organizing framework for this emerging area.
  • Survey goals: The organized discussion highlights key research trends and pinpoints open challenges to foster future research in conversational user simulation.The survey also directs readers to an appendix for details on related surveys and distinctions from this work.

2 Problem Definition

The paper defines conversations as temporally ordered sequences of speaker–utterance turns among heterogeneous participants and formulates simulation as generating a target participant’s next utterance from dialogue history and participant context.

  • Conversation representation: Participants may be human users, systems, or other agents, enabling user–user, user–system, and multi-party conversation simulations.The participant set is P = {p1, p2, . . . , pN}, containing N participants that can differ in type.
  • Conversation representation: A conversation is a temporally ordered sequence of T turns, each pairing a speaking participant with an utterance from vocabulary V.The formal representation is C = (c1, c2, ..., cT), with each turn containing the speaker and their produced utterance.
  • Simulation objective: Conversational simulation models one or more target participants by generating a participant’s next utterance from conversational history and participant-specific context.Participant context may include demographics, domain knowledge, or personal context.

3 Who: Simulated Users and Interactions

The section organizes conversational user simulation into four levels of granularity: general users, personas, roles, and individuals. It also notes that these paradigms often overlap in practice, while systematic hybrid simulation remains limited and relevant to multi-agent interactions.

  • Simulation Granularity: The taxonomy progresses from general-user simulation to persona-level, role-play, and individual-user conversation simulation.These levels move from broad population modeling toward increasingly fine-grained conversational targets.
  • General User Simulation: General-user simulation models conversations from a broad population perspective without providing distinguishing characteristics in the prompt.The default persona represents an average user sampled randomly from the general population.
  • Persona-Level User Simulation: Persona-level simulation goes beyond general users by modeling demographic, interest, or stylistic attributes explicitly.Persona grounding methods include demographic prompting, psychometric modeling, trait-infused architectures, and activation-level control, with fairness and bias concerns.
  • Role Play Simulation: Role-play simulation models real or fictional individuals with greater flexibility than fixed-trait personas, using identity handles and LLM knowledge of character traits.Prompting, finetuning, and self-play support applications in storytelling, social simulation, and agent memory modeling.
  • Individual User Simulation: Individual-user simulation is the taxonomy’s most fine-grained level, grounding behavior in users’ explicit and often dynamic personal contexts.Personal context can include chat logs, documents, interaction histories, dialogue histories, multi-session memory, and real-world traits.
  • Hybrid User Simulation: Although the four paradigms are conceptually distinct, large-scale models often blend them, and systematic hybrid user simulation remains limited.Hybrid simulation is particularly relevant to multi-agent interactions because emergent behavior depends on balancing generic user roles with individual variation.

4 What: Simulation Objectives

This section categorizes LLM-based user simulation by interaction pattern into Human–AI, Human–Human, AI–AI, and Many-Human–AI paradigms. It also highlights hybrid simulation as an under-theorized mixture of these paradigms requiring systematic modeling and benchmarking.

  • Taxonomy: The taxonomy defines four simulation paradigms by interaction pattern: Human–AI, Human–Human, AI–AI, and Many-Human–AI.These paradigms are defined over conversational trajectories consisting of turn-based utterance–response pairs.
  • Human–AI Simulation: Human–AI simulation pairs human utterances with fixed-model AI responses to mimic behavior and create realistic contexts for evaluating AI capabilities.Recent methods generate synthetic dialogues for annotation reduction, broad coverage, domain adaptation, and task-specific applications.
  • Human–Human Simulation: Human–Human simulation models two profiled participants to replicate natural dialogue, maintain consistent personas, and support grounded interactions.PersonaChat, Wizard-of-Wikipedia, EmpatheticDialogues, and MultiWOZ are key datasets, while self-play bootstrapping offers a scalable alternative to manual curation.
  • AI–AI Simulation: AI–AI simulation uses two autonomous agents that converse from a general seed prompt without ongoing human input, enabling scalable data generation and emergent-behavior study.Research examines emergent social behaviors, collaborative task-solving, and adversarial debate; entity-level personas, roles, and individual profiles are not specified.
  • Many-Human–AI Simulation: Many-Human–AI simulation extends Human–AI interaction to multiple human participants and one or more AI agents, capturing individual behavior and group dynamics.Although recent work explores collaborative roles, proxy participation, and group dialogue, general-purpose frameworks for this setting remain limited.
  • Hybrid Simulation: Hybrid simulation combines Human–Human, AI–AI, and Human–AI interactions in one environment, but remains under-theorized and needs systematic modeling and benchmarking.Smallville illustrates this mixture by simulating AI–AI communities whose individual dialogues resemble Human–Human exchanges.

5 How: Techniques and Methodologies

Section 5 organizes conversational user simulation techniques into prompting, retrieval-augmented generation, fine-tuning, reinforcement or preference optimization, and hybrid pipelines. These methods incorporate contextual knowledge and target realism, coherence, personalization, controllability, efficiency, and strategic behavior through different training and conditioning mechanisms.

  • Prompt-based simulation: Prompting follows method-driven and content-driven tracks, using zero-/few-shot or chain-of-thought methods alongside persona and task-specific prompts.Zero-/few-shot prompts support scalable simulation with minimal examples, chain-of-thought improves coherence through step-by-step reasoning, and content-driven prompts guide tone and domain behavior.
  • Retrieval-augmented generation: RAG conditions simulated responses on external knowledge through always-on, adaptive, or goal/state-driven retrieval triggers.Always-on methods retrieve passages every turn, adaptive methods use a learned classifier to decide when retrieval is needed, and goal/state-driven methods retrieve from internal user memory.
  • Fine-tuning: Fine-tuning trains user simulators on dialogue context, persona descriptions, optional task instructions, and ground-truth utterances using supervised objectives.The survey distinguishes full-model supervised training, parameter-efficient adapters or activation steering, and interactive optimization through interaction feedback.
  • Reinforcement and preference optimization: RLHF and DPO optimize user policies with rewards or preference scores across multi-turn interactions, enabling adaptive, strategic, and goal-driven behavior.Examples target latent-trait disambiguation, memory selection for factual coherence, and long-horizon planning through hierarchical utterance- and token-level control.
  • Hybrid approaches: Hybrid pipelines combine prompting, retrieval, fine-tuning, and RL/DPO to improve realism, controllability, grounding, personalization, and sample efficiency.Examples include retrieval-augmented fine-tuning, prompt-to-fine-tune bootstrapping, RAG plus RL/DPO loops, hierarchical pipelines, and personalized stacks.

6 Evaluation

Conversational user simulation is evaluated through traditional metrics, human assessment, and LLM-based judging, with recent work also emphasizing trustworthy and causal evaluation.

  • Evaluation methods: Human evaluation remains the gold standard, typically conducted through interactive or offline evaluation.Traditional metrics such as BLEU, ROUGE, and slot-F1 remain common but capture narrow facets and are often complemented by human or LLM judges.
  • Evaluation methods: Traditional metrics including BLEU, ROUGE, and slot-F1 provide efficient and reproducible evaluation for structured or goal-oriented dialogues.These metrics capture narrow facets of conversational quality.
  • Evaluation methods: LLM-as-Judge prompts a strong language model with dialogue context, generated responses, and rubrics targeting coherence, factuality, or safety.Typical protocols define evaluation dimensions and rating scales, such as 1–5 Likert scales, and provide few-shot exemplars or calibration prompts.
  • Evaluation methods: Recent studies emphasize trustworthy and causal/offline evaluation to assess reliability, robustness, and generalization beyond output accuracy.These paradigms examine uncertainty, distribution shifts, adversarial prompts, and generalization across topics and user profiles.

7 Datasets

The survey categorizes commonly used user-simulation datasets across dialogue types and interaction settings. These include personalized, multiparty, information-seeking, role-based, character-grounded, and negotiation conversations, supporting persona modeling, goal-driven interaction, and strategic reasoning.

  • Dataset categorization: User-simulation datasets are categorized across dialogue types, with a summary in Table 11 and additional details in Appendix F.The datasets span a wide range of interaction settings.
  • Interaction settings: Datasets cover personalized conversations, multiparty social dialogues, and information-seeking question answering.These settings represent different conversational interaction types.
  • Simulation objectives: Role-based and character-grounded simulations support fine-grained persona modeling, while negotiation datasets evaluate goal-driven interaction and strategic reasoning.The datasets extend beyond general dialogue to persona- and goal-oriented simulations.

8 Applications

Conversational user simulation supports a broad range of applications, including data augmentation, conversational recommendation, and educational conversational companions. LLMs have further benefited these applications by enabling adaptation to diverse user preferences and supporting interactive learning experiences.

  • Applications: User simulators support data augmentation where user histories are sparse, and LLMs have further benefited these applications.The passage attributes early data-augmentation use to Zhao et al. (2021) and later benefits from LLMs to Zhao et al. (2024a).
  • Applications: In conversational recommendation, simulators help systems adapt to diverse user preferences.This application is associated with Yoon et al. (2024).
  • Applications: In education, simulators power conversational companions that support interactive learning experiences.The supplied passage introduces this educational application but ends before specifying further details.

9 Open Problems & Challenges

LLM-based conversational user simulation still struggles with consistency, realism, diversity, and safety. Progress requires stronger memory, planning, controllable simulation, multi-user modeling, and robust quality and safety protocols.

  • Consistency and Long-Horizon Interaction: Extended interactions cause persona drift in style, beliefs, and goals, while broken memory and contradictions can trigger hallucinations and character violations.These problems are amplified in role-based settings.
  • Consistency and Long-Horizon Interaction: Simulated users may behave unrealistically cooperatively, accumulate errors, or lose task focus during long dialogues.The passage identifies better memory mechanisms, discourse planning, and consistency modeling as needed solutions.
  • Diversity and Interaction Realism: Simulators often produce overly polite, culturally homogeneous behavior, requiring finer control over emotion, verbosity, strategy, and other traits.Prompting enables personas, but diversity remains limited.
  • Diversity and Interaction Realism: Most research targets single-user setups and neglects hybrid or multi-user dynamics needed for realistic and personalized simulation.The passage presents multi-user modeling as an important realism and personalization challenge.
  • Safety and Evaluation: LLM-based simulations risk encoding bias and generating toxic content, while robust protocols for simulation quality and safety remain lacking.Risks are especially salient for personas involving sensitive demographics or public figures.

10 Conclusion

The survey organizes LLM-based conversational user simulation through a unified Who–What–How framework, reviews methods and evaluations, and identifies key open challenges.

  • 10 Conclusion: The survey structures representative literature along three axes: Who, What, and How.It uses this unified framework to organize research on LLM-based conversational user simulation.
  • 10 Conclusion: It reviews simulation methods, discusses their strengths and limitations, and categorizes them across diverse applications.The survey also examines evaluation protocols and common datasets to support benchmarking.
  • 10 Conclusion: The survey concludes by outlining key open challenges for future research.

11 Limitations

The survey’s taxonomy may not fully accommodate hybrid or domain-specific methods, and the work does not provide a comprehensive benchmark across methods.

  • The taxonomy balances generality and clarity but may not perfectly accommodate certain hybrid or domain-specific methods.These methods may fall outside the proposed categories.
  • A full benchmarking study comparing methods was beyond the scope of the survey.The survey summarizes datasets and evaluations without conducting comprehensive cross-method benchmarking.
  • The survey provides an overview of LLM-based conversational user simulation but acknowledges several limitations.The stated limitations include taxonomy fit and the absence of full benchmarking.

12 Ethical considerations … B.3 Role Play Simulation

The survey situates LLM-based conversational user simulation within broader simulation and human-modeling research, organizes existing surveys and methods by simulation target, and examines ethical risks. It covers general, persona-level, and role-play simulation, including their techniques, applications, and limitations.

  • A.1 Simulation in Other Fields: Simulation spans scientific disciplines, computer science, reinforcement learning, recommender systems, and counterfactual reasoning, motivating conversational user simulation.Examples include physical-system simulation, click models, handcrafted RL environments, human-centric simulators, and treatment-effect estimation from logged data.
  • A.2 Human Modeling: Human preferences are modeled with Bradley-Terry-Luce and Plackett-Luce discrete choice models, while human-feedback learning has applications in machine learning.The passage identifies pairwise and permutation preferences as established modeling targets and connects them to reinforcement learning from human feedback.
  • A.3 Comparing to Existing Surveys: Compared with related surveys, this work analyzes varied simulation targets comprehensively rather than focusing on one target or only on conversational data generation.The comparison explicitly distinguishes the survey from work focused on a simple simulation target and from surveys limited to data generation.
  • B.1 General User Simulation: General user simulation improves dialogue systems through turn-level and trajectory-level optimization, including preference learning, exploration, selective segment use, and policy refinement.Methods described include M-DPO, ETO, SDPO, AgentQ’s MCTS with self-critique, and LOOP’s environment-based PPO variant.
  • B.2 Persona-level User Simulation: Persona-level simulation assigns explicit characteristics through prompting, psychosocial frameworks, personal context, fine-tuning, preference optimization, or activation steering.These approaches aim to produce realistic and consistent psychometric traits and representative linguistic patterns.
  • B.2 Persona-level User Simulation: Persona assignment can create systematic bias, skew political forecasts, increase toxicity, and reinforce harmful stereotypes, requiring stronger validation and trustworthy practices.The passage identifies bias as a critical limitation of persona-level simulation rather than merely a general ethical concern.
  • B.3 Role Play Simulation: Role-play simulation exploits LLMs’ implicit character knowledge to generate diverse, persona-driven behavior and decisions, including complex social interactions and interactive human-behavior simulacra.Character-level benchmarks evaluate decision-making consistency, while emerging work explores characters’ autobiographical knowledge and lifelong learning.
  • B.3 Role Play Simulation: Role-play methods use prompting, character-dialogue fine-tuning, adversarial self-play, and masked-dialogue reconstruction to improve output style, naturalness, consistency, and quality.These techniques are intended to elicit or shape implicit character knowledge and reduce deviations from assigned characters.

B.3.1 Evaluation on Role-Playing Benchmarks … H.3 Video Understanding

The survey organizes conversational user simulation by role-playing performance, user granularity, simulation objectives, techniques, evaluation, applications, and challenges. Across these areas, it highlights the strengths and limitations of prompting, retrieval, fine-tuning, interactive learning, human assessment, and persona-grounded simulation.

  • B.3.1 Evaluation on Role-Playing Benchmarks: Fine-tuned models often improve role-specific knowledge and character consistency over prompt-based or retrieval-augmented alternatives on role-playing benchmarks.Prompt-based models benefit from scale and implicit knowledge, whereas fine-tuned models—often using smaller backbones—show stronger SPE performance; ChatPLUG generally underperforms.
  • B.4 Individual User Simulation: Individual user simulation grounds responses in explicit personas, dialogue history, multi-session memory, real personality traits, and naturally occurring interactions.PERSONACHAT and FoCus use profiles and facts; PHMN models communication patterns; MSC and Mem0 support cross-session memory; RealPersonaChat and LiveChat capture personality and real-world variation.
  • C What: Simulation Objectives (Extended): Simulation objectives span human–AI data generation, human–human dialogue construction, AI–AI cooperation and debate, and many-human collaboration.These settings address data scarcity, task-oriented training, emergent social dynamics, multi-agent problem solving, and team participation, while many-human–AI simulation remains underexplored.
  • G Applications (Extended): Applications include medical and museum dialogue generation, recommendation, education, limited-feedback settings, HCI/UI studies, and video understanding, but simulated conversations can remain distinguishable from human data.Hämäläinen et al. emphasize careful prompting and early design use, while the supplied challenge passage identifies persona consistency versus adaptability as an unresolved issue in long conversations.
  • D How: Techniques and Methodologies (Extended): Prompting, RAG, fine-tuning, RL/DPO, and hybrid pipelines provide complementary mechanisms for controlling behavior, grounding responses, adapting policies, and improving long-horizon coherence.Prompting uses exemplars, reasoning, personas, and task constraints; RAG varies retrieval triggers; fine-tuning ranges from full-model updates to adapters and self-play; RL/DPO targets personalization, memory, planning, hindsight learning, and clarification.
  • D.5 Hybrid Approaches: Hybrid methods combine retrieval, prompting, fine-tuning, and RL/DPO to balance adaptability, evidence use, memory, persona stability, and strategic dialogue behavior.Prompt-to-fine-tune curricula bootstrap synthetic conversations before supervised adaptation, while other systems couple retrieval with preference optimization or organize specialized planner, retriever, executor, and critic modules.
  • E Evaluation (Extended): Traditional metrics offer inexpensive, reproducible signals but capture narrow dimensions of quality, motivating complementary human or LLM evaluation.N-gram overlap correlates poorly with open-ended human judgments, slot accuracy omits pragmatic qualities, perplexity does not necessarily reflect interactive preference, and task success requires a defined objective.
  • H.3 Video Understanding: A central challenge is balancing stable persona expression with adaptation to evolving context, because excessive consistency can cause rigidity while excessive adaptability can cause persona drift.This trade-off remains largely unexplored in multi-turn and open-ended interactions.
Loading 2604.24977v1…