PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Single-Turn: A Survey on Multi-Turn Interactions with Large Language Models

Yubo Li, Xiaobin Shen, Yidi Miao, Xinyu Yao, Xueying Ding, Ramayya Krishnan, Rema Padman

arXiv:2504.04717v6cs.CLcs.AI

TL;DR

Single-turn benchmarks do not fully capture the stateful, context-dependent demands of sustained LLM interaction. This survey responds with a task-oriented review of multi-turn evaluation and enhancement methods, finding that progress increasingly depends on managing interaction history and auxiliary state rather than model scale alone. It also identifies persistent limitations in knowledge adaptation and sociolinguistic coverage.

  • Problem

    Single-turn evaluation underexplores real-world multi-turn interaction, where models must maintain context, consistency, user-intent alignment, and responsiveness across sequential dialogue.

  • Method

    The survey uses a task-oriented taxonomy and reviews benchmarks, model-centric methods, external integrations, agent-based approaches, and open challenges.

  • Results

    The survey concludes that improvement increasingly comes from better management of interaction history and auxiliary state rather than model scale alone.

  • Takeaways & Limitations

    Future progress requires stronger long-horizon context management, more reliable adaptation, and more auditable evaluation for high-stakes interaction settings.

  • Takeaways & Limitations

    On-the-fly learning from user-provided information remains difficult because of catastrophic forgetting and risks from false or malicious inputs.

Abstract

from arXiv · show

Recent advances in large language models (LLMs) have substantially improved single-turn task performance, yet real-world applications increasingly demand sophisticated multi-turn interactions. This survey provides a comprehensive review of recent progress in evaluating and enhancing multi-turn LLM interactions. Centered on a task-oriented taxonomy-spanning instruction following in domains such as mathematics and coding, and conversational engagement in role-playing, healthcare, education, and adversarial jailbreak settings-we systematically examine the challenges of maintaining context, coherence, fairness, and responsiveness across prolonged dialogues. We organize existing benchmarks and datasets into coherent categories reflecting the evolving landscape of multi-turn dialogue evaluation, and review a broad spectrum of enhancement methodologies, including model-centric strategies (in-context learning, supervised fine-tuning, reinforcement learning, and architectural innovations), external integration approaches (memory augmentation, retrieval-based methods, and knowledge graphs), and agent-based techniques for collaborative interaction. Finally, we identify open challenges and promising directions for future research to further improve the robustness and effectiveness of multi-turn LLM interactions.

1 Introduction

This survey addresses the gap between strong single-turn evaluation and the context-dependent demands of real-world multi-turn LLM interaction. It organizes the field by task, reviews methods for improving sequential dialogue, and identifies open challenges.

  • Motivation: Single-turn evaluation underexplores multi-turn interaction, where models must sustain context, maintain consistency, handle ambiguity, and respond across sequential turns.Multi-turn dialogue also introduces dialogue coherence, user-intent alignment, cumulative errors, hallucinations, and contextual drift.
  • Scope: The survey covers sequential settings where LLMs maintain context, adapt across turns, and achieve task success over dialogue trajectories.Its scope includes instruction following and conversational engagement across high-impact domains.
  • Taxonomy: The taxonomy spans mixed-topic instruction following and open-ended conversational engagement across several application domains.The survey emphasizes practical scenarios and characteristics of multi-turn deployments.
  • Improvement Methods: Improvement methods are grouped into model-centric, external-integration, and agent-based approaches.These include in-context learning, fine-tuning, reinforcement learning, architectural innovations, memory, retrieval, knowledge graphs, and collaborative agents.
  • Open Challenges: The survey organizes open challenges into Context Understanding, Complex Reasoning, Adaptation & Learning, Evaluations, and Ethical & Safety Issues.It uses these areas to frame future research directions for sustained interaction.
  • Positioning: The survey claims a task-oriented taxonomy jointly covering instruction following and conversational engagement with benchmarks, improvement methods, and open challenges across seven application subdomains.This scope distinguishes the survey from prior work organized primarily around single-turn evaluation or methodological instruction tuning.

2 Background and Problem Formulation

The paper models multi-turn interaction as a stateful trajectory in which inputs, responses, history, latent state, and environment feedback evolve across turns. It distinguishes static offline dialogue evaluation from interactive policy evaluation, while treating richer agentic environments as adjacent scope.

  • 2.1 A Sequential View of Multi-Turn Interaction: A multi-turn interaction is represented as a trajectory of alternating queries and responses, τ = (q1, r1, q2, r2, . . . , qT, rT).The query may include user input, dialogue history, retrieved passages, tool outputs, or external memory.
  • 2.1 A Sequential View of Multi-Turn Interaction: At turn t, the model conditions on interaction history and may use auxiliary state such as retrieved evidence, memory slots, or structured task state.Finite context windows can require truncation, summarization, retrieval, or external memory for long interactions.
  • 2.1 A Sequential View of Multi-Turn Interaction: Multi-turn quality depends on how each response shapes future turns, not only on its local helpfulness.The trajectory-level utility may represent instruction satisfaction, factual correctness, pedagogical quality, safety, user satisfaction, or downstream task success.
  • 2.2 Interactive Tasks vs. Static Dialogue Evaluation: Static evaluation uses fixed histories and targets, whereas interactive evaluation depends on the model response and assesses the induced trajectory.Interactive performance can be measured by trajectory-level success, cumulative utility, or final-task completion.
  • 2.2 Interactive Tasks vs. Static Dialogue Evaluation: Dynamic agent settings involving planning, tools, environment manipulation, or multi-agent coordination are treated as adjacent rather than core scope.They remain relevant when discussing improvement methods, but the survey does not aim to cover agent literature exhaustively.
  • 2.2 Interactive Tasks vs. Static Dialogue Evaluation: An interactive task is one where turn-level model decisions causally affect later observations, including when the environment is a human or user simulator.This distinguishes policy-dependent interaction from fixed dialogue logs.
  • 2.2 Interactive Tasks vs. Static Dialogue Evaluation: The paper distinguishes multi-turn interaction from independent single-turn prompts that merely share a topic.It defines memory broadly as mechanisms that preserve or reconstruct relevant information beyond immediate local generation state.

3 Multi-Turn Interaction Tasks

The survey categorizes multi-turn interactions by task rather than isolated capabilities, distinguishing instruction following from conversational engagement according to user-intention clarity and task complexity.

  • The taxonomy organizes multi-turn interactions by concrete tasks because reasoning, memory, contextual understanding, and adaptability operate together toward user goals.
  • Task families are distinguished primarily by user-intention clarity and task complexity.
  • The survey’s taxonomy is task-oriented and summarized in Figure 2, with fuller literature coverage in Appendix Table 11.

3.1 Instruction Following Tasks

Multi-turn instruction-following research spans general, mathematical, and coding tasks, with benchmarks increasingly testing longer dialogues, structural constraints, fairness, feedback responsiveness, and generalization. Evaluation resources and methods have expanded, but automated judging remains limited by fairness, contextual sensitivity, reference dependence, and insufficient human validation.

  • 3.1 Instruction Following Tasks: Benchmarks cover general instruction following, mathematics, and coding, reflecting diverse multi-turn task demands.
  • 3.1.1 Instruction Following Tasks in General: MT-Bench introduced 80 two-turn dialogues across eight categories and evaluated responses through pairwise comparisons using LLM judges.
  • 3.1.1 Instruction Following Tasks in General: MT-Bench++ extended dialogues to eight turns, while MT-Bench-101 assessed perceptivity, interactivity, and adaptability across 4,208 turns and 1,388 dialogues.
  • 3.1.1 Instruction Following Tasks in General: MT-Eval found significant multi-turn performance degradation, while newer benchmarks test multilingual constraints, fairness, feedback handling, consistency, and interactive planning.
  • 3.1.2 Instruction Following Tasks in Math: Math studies report weak generalization to new problems, while supervised fine-tuning and preference-based reinforcement learning improve downstream performance.
  • 3.1.4 Discussions: Dataset sizes grew from 80 examples in MT-Bench to over 1,000 in newer benchmarks, alongside increasingly automated and LLM-assisted curation.
  • 3.1.4 Discussions: LLM-as-a-judge methods offer scalable evaluation, but preference leakage, contextual sensitivity, reference dependence, and limited human agreement checks threaten reliability and fairness.

3.2 Conversational Engagement

Conversational-engagement research evaluates sustained multi-turn performance across general and specialized domains, while developing persona, retrieval, fine-tuning, and healthcare-focused methods. The literature also exposes weaknesses in automated evaluation and adversarial robustness.

  • Overview: Conversational-engagement benchmarks span role-play, healthcare, education, and jailbreak settings, alongside general frameworks for sustained interaction quality.ABC-Eval measures 16 turn-level conversational behaviors, while later work examines human-likeness, naturalness, and simulator reliability.
  • Overview: BotChat reduces reliance on costly human judges by using LLMs to generate extended conversations from real-world dialogue seeds and evaluate them automatically.GPT-4 serves as an automated judge in the described paradigm.
  • Role-Play: Retrieval-based methods are especially effective for on-the-fly personalization, while specialized training can improve consistency for culturally specific or language-specific roles.The survey contrasts retrieval-based personalization with fine-tuned local models and global models across character benchmarks.
  • Role-Play: Role-play research targets persona consistency, character fidelity, and temporal consistency through prompting, persona-conditioned training, and specialized benchmarks.CharacterEval, RoleEval, and TimeChara examine behavioral realism, character knowledge, and leakage of future story events.
  • Healthcare: Healthcare systems combine dialogue-based knowledge encoding, term-aware training, and reasoning layers, but adversarial evaluation shows that standard benchmark performance does not ensure real-world reliability.MedFuzz introduces ambiguous patient traits and biased data to expose clinical weaknesses.

Intelligent Tutoring Systems

Intelligent tutoring research shifts LLMs from answer delivery toward Socratic, strategy-guided, adaptive, and context-aware multi-turn instruction. Benchmarks and classroom-oriented simulations evaluate both pedagogical behavior and learner support.

  • Socratic and Strategy-Guided Tutoring: Socratic and guided tutors use open-ended questions, planned question hierarchies, and teaching strategies to prompt student reasoning instead of directly revealing solutions.SocraticLM uses 35k simulated teacher–student dialogues, while TreeInstruct plans question hierarchies for debugging code.
  • Adaptive Tutoring: Adaptive tutoring systems model learner styles, personas, and evolving reasoning so explanations and assistance can be tailored during multi-turn exchanges.PACE adapts explanations to learning styles, and JeepyTA provides context-specific help in an online course forum.
  • Evaluating LLM Tutoring in Mathematics: MathDial evaluates tutoring with simulated misconceptions, expert annotations, student success, telling@k, and measures of misconception correction and teacher response quality.Fine-tuning on MathDial improved student outcomes by emphasizing hints and prompting over premature answers.
  • Socratic and Strategy-Guided Tutoring: Retrieval augmentation and alignment methods ground tutor responses in course materials while targeting safer, more pedagogically appropriate, and query-relevant outputs.CourseAssist uses syllabi and lecture notes to align generated responses with instructor expectations.
  • Bridging Research and Real Classrooms: Recent benchmarks make classroom evaluation more concrete through dialogue-based, verifiable, and rubric-backed assessments spanning explanation, feedback, and pedagogical dimensions.KMP-Bench, MRBench, and TutorBench extend evaluation beyond subject correctness alone.

Automated Feedback & Grading Support

Automated feedback and grading research seeks scalable educational support but remains constrained by faithfulness, pedagogical alignment, rubric consistency, and human-level nuance. Multi-turn simulations provide a complementary sandbox for testing educational interventions before classroom deployment.

  • Automated Feedback Support: LLM-generated feedback can sound insightful while adding content not grounded in student work, making faithfulness and alignment central research problems.Reported failures include fabricated critiques and irrelevant suggestions that may confuse learners.
  • Automated Feedback Support: Verification, rubric-based reinforcement learning, and simulated student revisions target more accurate, misconception-aware, and pedagogically aligned feedback.These methods analyze reasoning, encode preferences such as correctness and encouragement, and close the feedback loop through revisions.
  • Automated Feedback Support: Classroom deployments find automated comments broadly helpful, but instructors report that outputs do not always reflect assignment-specific pedagogical objectives.This creates a gap between perceived usefulness and alignment with instructional goals.
  • Automated Grading Support: LLM graders can approach human-level consistency and accuracy for some assignments, yet essay scoring still struggles with consistent rubric application and nuanced feedback.A study of 463 Master’s-level exam responses found about 70% of AI scores within 10% of human scores and 31% within 5%.
  • Scenario Simulation: Multi-turn educational simulations let researchers prototype interventions rapidly, ethically, and inexpensively before involving real students.SimClass also demonstrates that multiple LLM agents can model teacher–student and peer-like classroom interactions that support engagement.

4 Improvements

The survey organizes improvement methods into model-centric, external-integration, and agent-based approaches, while noting that papers may span branches. Evidence across prompting, fine-tuning, reinforcement learning, and architectural methods shows that multi-turn gains are strongly task- and strategy-dependent.

  • Improvement taxonomy: The taxonomy groups methods into model-centric, external-integration, and agent-based approaches, while assigning multi-branch papers by their primary contribution.These methods respectively refine the model, add external resources, or support iterative action, reflection, and collaboration.
  • In-Context Learning: Naive multi-turn in-context prompting can degrade sequential-dialog performance, whereas interactive reasoning and plan-and-refine strategies can improve code generation using compiler feedback.Prompting outcomes vary across tasks: persona prompts may not improve factual accuracy, while structured personas and memory can support personalized dialogue.
  • In-Context Learning: Explicit state and graph representations improve long-distance constraint tracking and multi-hop performance without requiring additional model training.Graph-based prompts and selective state reminders replace raw-history prompting with structured context representations and can reduce token costs.
  • Supervised Fine-Tuning: Multi-turn supervised fine-tuning benefits from domain-specific interaction data and complementary improvements in dialogue-data generation and fine-tuning strategy.The survey links these approaches to better context retention and more coherent multi-turn responses.
  • Reinforcement Learning: Reinforcement-learning results are mixed: specialized medical and coding approaches can yield substantial gains, while generalized RLHF often has limited or negative multi-turn effects.Hierarchical methods such as ArCHer and SCoRe address multi-turn credit assignment and self-correction through structured training and reward design.
  • Reinforcement Learning: 77.5% to 83.9%: a 7B model’s GSM8K accuracy after multi-turn preference optimization.This result illustrates the reported gains from optimizing models for multi-turn preferences rather than isolated responses.
  • Architectural Innovations: Memory and recurrence mechanisms capture long-term dependencies across segments, improving multi-turn dialogue coherence, performance, and efficiency.These architectural mechanisms are presented as another route to more robust and scalable conversational systems.

4.2 External Integration Approaches

External integration methods augment LLMs with memory, retrieval, and knowledge graphs to improve context use, factuality, and reasoning across extended interactions. The survey also describes agent-based systems that use iterative action, planning, and collaboration, while noting coordination and verification risks.

  • External Integration: External integration augments LLMs with memory, retrieval, and knowledge graphs to improve information retrieval, verification, reasoning, reliability, accuracy, and consistency.These resources are intended to mitigate compounding errors and misinformation propagation in extended interactions.
  • Memory Augmentation: Memory-augmented methods store and recall past interactions to correct misinterpretations, reduce repeated errors, and adapt to evolving dialogue.MemPrompt retrieves earlier user corrections, while newer designs optimize indexing, retrieval, reading, episodic storage, and hierarchical organization.
  • Memory Augmentation: External, episodic, and hierarchical memory mechanisms can enhance consistency and contextual understanding in multi-turn conversations.The surveyed systems include prospective and retrospective memory, hypergraph-based history organization, and boundary-aware retrieval.
  • Retrieval-Augmented Generation: Retrieval-augmented generation conditions responses on relevant external documents to improve factuality and informativeness, with Wizard of Wikipedia reporting higher factual accuracy and user engagement than non-retrieval baselines.Multi-turn RAG evaluation includes MTRAG, CORAL, and RAD-Bench, which test retrieval, generation, citation labeling, and reasoning across conversational settings.
  • Retrieval-Augmented Generation: RAGate learns when retrieval helps because indiscriminate retrieval can increase uncertainty and hallucination risk.Recent RAG work therefore considers conversational history and turn-level retrieval decisions rather than augmenting every system turn automatically.
  • Knowledge Graphs: Knowledge graphs and graph neural networks support entity tracking, coreference resolution, dialogue-context management, and structured reasoning.Some methods integrate graph-derived embeddings during pre-training or fine-tuning, while others construct graphs from historical dialogue to enrich context.
  • Agent-Based Approaches: Single-agent systems use iterative interaction with tools, environments, and internal memory to improve reasoning, decision-making, and self-improvement across multi-turn settings.Voyager exemplifies lifelong learning in an open-ended virtual environment, while other systems report gains on coding and sequential decision-making benchmarks.

5 Open Challenges

Multi-turn LLMs face interconnected challenges in retaining evolving context, resolving ambiguity, reasoning across turns, adapting safely, and being evaluated fairly and reliably. These difficulties span multilingual interaction, persistent memory, adversarial manipulation, sociolinguistic diversity, and bias accumulation.

  • Context Understanding: Models struggle to track distributed, evolving conversational state, including implicit references, user attributes, and temporally changing information.Failures include incorrect antecedent resolution, forgotten or conflated user facts, and degraded performance when instructions are distributed across turns.
  • Context Understanding: Clarification remains an open challenge: models often answer underspecified queries instead of initiating grounding repairs or targeted follow-up questions.Benchmarks identify under-clarification across missing-information, overconfidence, and ambiguity settings, while payoff-aware question selection and proactive information gathering improve performance.
  • Complex Reasoning: Early errors and false premises can persist and amplify across dialogue trajectories, producing compounding reasoning failures, especially in sequential decision tasks.General-purpose models are also largely reactive, lacking robust policies for deciding when to ask versus when to inform in interactive domains.
  • Adaptation & Learning: Persistent knowledge adaptation remains unsolved because systems must update from user-provided information while balancing plasticity, stability, catastrophic forgetting, and malicious data injection.External memory and dynamic context augmentation provide surrogates, but raw history consumption remains brittle and retrieval depends on pre-established knowledge sources.
  • Ethical & Safety Issues: Multi-turn safety is constrained by staged manipulation, misinformation propagation, prompt leakage, and abrupt safety collapse as models adapt to harmful interaction histories.Studies report policy violations and degraded fraud resistance under gradual credibility building, urgency creation, and emotional manipulation.
  • Evaluations: Evaluation remains limited by Western and English-centric coverage, accumulated bias, and judge reliability that can drift across adversarial rounds and victim models.Fairness metrics across diverse sociolinguistic profiles remain largely unsolved, while scalable evaluation requires auditable, bias-aware, multi-perspective judging.

6 Conclusion

The survey frames multi-turn interaction as a distinct paradigm requiring coordinated capabilities across sustained dialogues and presents a task-oriented synthesis of methods, applications, and unresolved challenges.

  • Taxonomy: The task-oriented taxonomy organizes multi-turn interactions around instruction following and conversational engagement rather than isolated capabilities.It covers application domains including role-play, healthcare consultation, education, and jailbreak settings.
  • Multi-turn interaction: Multi-turn use more closely reflects real-world interaction and requires context retention, coherent cross-turn reasoning, adaptive behavior, and handling of evolving intent.
  • Improvement methods: The surveyed improvement methods span model-centric techniques, external memory and retrieval strategies, knowledge graphs, and agent-based frameworks.The review includes in-context learning, fine-tuning, reinforcement learning, architectural innovations, and single- or multiagent systems.
  • Open challenges: Extended conversations still suffer from coherence failures, cross-turn reasoning errors, limited adaptation, and unevenly reported evaluation conditions.The survey organizes these unresolved issues into a roadmap for future research.
  • Future directions: Future progress requires stronger long-horizon context management, more reliable adaptation, and more auditable evaluation for high-stakes interactions.

Broader Impact Statement

The survey’s broader impact is shaped by search, scope, taxonomy, and domain-emphasis choices that constrain which multi-turn AI work it represents and how it is organized.

  • Search and availability bias: Search and availability bias may underrepresent work from lower-resource regions, proprietary deployments, and communities using different terminology.
  • Scope bias: The survey excludes multimodal systems, centers non-multimodal interaction, and treats agentic systems as adjacent rather than central.These choices improve coherence and tractability but prevent the survey from being a complete account of all multi-turn AI systems.
  • Taxonomy choices: Taxonomy assignments are interpretive because papers may combine benchmarks, methods, and analyses or cross boundaries between dialogue, agents, and multimodal interaction.Alternative placements can therefore be defensible.
  • Domain emphasis: Emphasizing healthcare, education, role-play, and jailbreak settings may understate other influential but less benchmarked or publicly documented deployments.

A Review Methodology Details

The review retrospectively documents a task-oriented narrative-review workflow using PRISMA-ScR-oriented reporting, multiple search channels, explicit eligibility rules, and manual screening and charting.

  • Reporting framework: The survey was retrospectively audited after an original narrative review, with PRISMA-ScR used as the primary reporting guide and PRISMA 2020 as supporting guidance.No prospective protocol was registered, and full compliance with every reporting item is not claimed.
  • Search strategy: Corpus construction used venue browsing, scholarly keyword searches, GPT-assisted and manual discovery, section-by-section refreshes, and citation chasing.
  • Search strategy: A representative query combined an LLM term, a multi-turn term, and a task or domain term, then was specialized by section.The reported template includes terms for instruction following, coding, healthcare, education, role-play, and safety.
  • Eligibility: Included papers studied non-multimodal multi-turn LLM interaction and contributed relevant benchmarks, datasets, methods, analyses, or challenge discussions with sufficient technical detail.
  • Reporting coverage: Table 10 maps key PRISMA-ScR reporting expectations to the documentation provided in the survey.
  • Screening and verification: Screening and full-text verification were conducted manually, while retrospective auditing recovered stage-level counts for flow diagrams.The original review was not prospectively logged with reviewer-by-reviewer screening tallies.
  • Data charting: Data charting recorded bibliographic identity, task and subsection placement, paper role, benchmark details, evaluation setup, judge type, and concise content notes.

A.1 Original Corpus Construction (2022 to April 2025)

The original 2022–April 2025 corpus was reconstructed through a documented workflow covering venue and database searches, supplementary discovery, scope filtering, and eligibility decisions.

  • Original corpus workflow: The reconstructed process included identification from venue-centered and scholarly searches, GPT-assisted and manual additions, scope filtering, and eligibility decisions.
  • Identification sources: The figure distinguishes studies identified via databases and registers from those identified through other methods.
  • Original corpus workflow: Figure 5 documents the original corpus-construction workflow for the 2022–April 2025 survey phase.

A.2 Revision Update (May 2025 to April 2026)

The revision update extended the survey through April 2026 using the same inclusion, exclusion, and boundary-setting procedures, adding newly included studies to the Version 2 corpus.

  • The update continued repository maintenance and candidate-paper collection after the April 2025 snapshot under the same search and screening procedures.The accumulated material was consolidated after TMLR reviews into a structured revision update.
  • The update-stage workflow records identification of new studies through databases and registers.
  • Figure 6 documents the retrospective workflow used for the revision update from April 2025 through April 2026.The flow diagram follows the same inclusion, exclusion, and boundary-setting rules.

B.1 Scope Boundaries and Adjacent Survey Directions

The survey focuses on text-based multi-turn LLM interaction while treating agentic systems and other dialogue literatures as adjacent rather than exhaustive core coverage.

  • Core scope: The core scope covers sequential LLM interaction evaluated for context maintenance, adaptation across turns, and task success over dialogue trajectories.The literature is organized by task family across instruction following and conversational engagement domains.
  • Scope distinctions: The survey distinguishes static dialogue evaluation, interactive multi-turn tasks, and dynamic environment or agent settings.
  • Adjacent surveys: Broader dialogue-system, domain-specific, capability-oriented, and conversational-agent surveys are positioned as complementary adjacent directions.
  • Adjacent agentic systems: Agentic systems are reviewed when they inform multi-turn interaction, but are not treated as an exhaustive core task family.Their richer action spaces commonly involve tools, planning, environment manipulation, or multi-agent coordination.
  • Excluded multimodal literature: Multimodal multi-turn systems are excluded because they involve different input spaces, task structures, evaluation protocols, and a rapidly expanding benchmark ecosystem.They are identified as important context for future extensions but remain outside the adopted text-only scope.

C Full Task Taxonomy

The survey maps multi-turn LLM literature through a task-oriented taxonomy spanning instruction following and conversational engagement, with representative benchmark, method, system, and analysis studies.

  • Table 11 organizes the discussed literature across instruction following (IF) and conversational engagement (CE).The map covers benchmark, method, system, and analysis papers.
  • Instruction Following: Instruction-following categories include math-specific dialogue and interactive coding, represented by benchmarks, reasoning methods, debugging, and steering studies.
  • Conversational Engagement: Conversational-engagement categories include overview benchmarks, role-play, healthcare, education, and adversarial jailbreak settings.
  • The full task-oriented map is continued across multiple table pages and includes category-level representative-work listings.
  • Conversational Engagement: The taxonomy includes representative works spanning pedagogical agents, math tutoring, healthcare knowledge sources, proactive diagnosis, and multi-turn exploitation studies.
Loading 2504.04717v6…