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

Introduction

• Single-turn evaluation overlooks history, evolving state, feedback, and error compounding, while sustained dialogue requires interpreting prior exchanges, refining goals, and adapting responses.The survey presents multi-turn interaction as a temporal, stateful trajectory rather than an isolated prompt-response exchange.Source: §1 · pp. 1–2, Figure 1 · p. 2

Scope

• The core scope is non-multimodal, task-oriented sequential LLM dialogue evaluated for context maintenance, cross-turn adaptation, and trajectory-level task success; agentic work is adjacent unless directly informative.Works centered primarily on agent planning, environment control, or multimodal perception and action are treated as adjacent rather than core scope.Source: §1 · pp. 2–3

Related Surveys

• Prior surveys variously focus on static instruction following, instruction tuning, domain dialogue, or capability-oriented multi-turn work, while this review adopts an application-oriented task taxonomy (Zhou et al., 2023; Zeng et al., 2024b; Zhang et al., 2026; Yi et al., 2025; Valizadeh & Parde, 2022; Shi et al., 2024c; Zhang et al., 2025a).The review integrates task structure, benchmark resources, improvement methods, and open challenges in a unified view.Source: §1 · p. 3

Review Methodology

• The survey retrospectively documents its April 2025 corpus process and extends coverage through April 2026 using a PRISMA-ScR-inspired transparency protocol, not a fully prospective systematic review.Appendix A reports search sources, keyword families, screening stages, criteria, corpus counts, and assignment logic.Source: §1 · p. 3

Main Contributions

• Its task taxonomy covers mixed-topic instruction following and open-ended conversational engagement, jointly organized with harmonized benchmark tables across seven application subdomains.The survey emphasizes high-impact areas such as healthcare and education.Source: §1 · pp. 3–4
• The survey groups improvement methods into model-centric strategies, external integration, and agent-based approaches for extended interactions.These include in-context learning, supervised fine-tuning, reinforcement learning, and architectural innovations; memory, retrieval, and knowledge graphs; and proactive agents acting individually or collaboratively.Source: §1 · p. 4
• It organizes open challenges into Context Understanding, Complex Reasoning, Adaptation & Learning, Evaluations, and Ethical & Safety Issues, then discusses future directions.The survey also summarizes overarching insights in its conclusion.Source: §1 · p. 4

2.1 A Sequential View of Multi-Turn Interaction

• A multi-turn interaction is modeled as trajectory τ=(q1,r1,...,qT,rT), where q_t can include the current utterance, dialogue history, retrieved passages, tool outputs, or external-memory summaries.Source: §2.1 · p. 4
• Because context windows are finite, long histories may require truncation, summarization, retrieval, or external memory; optional auxiliary state z_t can encode retrieved evidence, memory slots, or structured task state.Source: §2.1 · p. 4
• Each response updates latent state summarizing goals, discourse, task progress, and external context, so turn quality depends on future effects as well as local helpfulness and trajectory utility.Source: §2.1 · p. 4

2.2 Interactive Tasks vs. Static Dialogue Evaluation

• Static evaluation fixes dialogue transcripts or next-turn targets and scores prediction or ranking from histories; MT-Bench-family benchmarks largely remain offline (Zheng et al., 2023; Bai et al., 2024).Source: §2.2 · p. 5
• In interactive tasks, responses change later inputs, so evaluation concerns induced trajectories and measures trajectory-level success, cumulative utility, or final-task completion.Clarifying questions can reveal constraints, tutoring can alter a student’s reply, and debugging suggestions can trigger compiler feedback.Source: §2.2 · p. 5
• Tool use, external-environment manipulation, and multi-agent coordination are adjacent, richer-dynamics settings rather than the survey’s core non-multimodal dialogue scope (Yao et al., 2023; Liu et al., 2024e).Source: §2.2 · p. 5
• A task is interactive when turn-level choices causally affect later observations, unlike independent single-turn prompts that merely share a topic.Source: §2.2 · pp. 5–6
• Memory preserves or reconstructs information; ICL conditions on prompt demonstrations without parameter updates (Brown et al., 2020; Dong et al., 2024a), while RAG adds retrieved evidence.Source: §2.2 · pp. 5–6

• Task families provide a practical framework for analyzing multi-turn interactions, where reasoning, memory, contextual understanding, and adaptability work together toward user goals.Source: §3 · p. 6
• Instruction-following tasks center on explicit instructions and clear intentions, with performance judged by how precisely models comply or execute them.Source: §3 · p. 6
• Conversational engagement tasks involve open-ended or evolving intentions and may require proactive information seeking, synthesis, goal inference, or external knowledge and tools.The model may act as a health consultant, teaching assistant, or customer service representative while sustaining contextually appropriate responses.Source: §3 · p. 6

3.1 Instruction Following Tasks

• Benchmarks cover general instruction following, mathematics, and coding, with general-purpose evaluations most common and specialized math and coding tasks testing distinct challenges.The survey organizes benchmarks into these three categories, as illustrated in Figure 2.Source: §3.1 · p. 7

3.1.1 Instruction Following Tasks in General

• General benchmarks progress from MT-Bench’s two-turn dialogues to longer or more fine-grained tests, while studies find uneven multi-turn performance and compounding errors.MT-Bench++ extends interactions to eight turns; MT-Bench-101 assesses interaction skills, and MT-Eval finds that errors worsen with successive turns and greater temporal distance from relevant context.Source: §3.1.1 · p. 7
• Multilingual and fairness benchmarks expose context-retention brittleness across languages and assess bias across dialogue stages and attributes.M2Lingual evaluates 12 languages; Multi-IF covers verifiable instructions in eight languages, while FairMT-Bench evaluates stereotype and toxicity biases across six attributes.Source: §3.1.1 · pp. 7–8
• Other general-task work evaluates feedback responsiveness, response consistency, ambiguity resolution, conversation structure, and instruction priority across turns.FB-Bench distinguishes error correction from response maintenance; CARG targets consistency, while newer benchmarks and methods address clarification, proactive information gathering, instruction hierarchies, and task-agnostic reasoning or web navigation.Source: §3.1.1 · pp. 8–9

3.1.2 Instruction Following Tasks in Math

• Multi-turn math work combines dialogue, tools, and feedback: MathChat-Agent improves accuracy by roughly 6% over standard single-turn chain-of-thought prompts, while MINT reports 1–17% accuracy gains per tool-use or feedback turn.MathChat evaluations show sharp performance drops on sustained multi-turn tasks, while fine-tuning on MathChatSync improves multi-turn performance; MINT also finds that single-turn-only fine-tuning can underperform in multi-turn settings.Source: §3.1.2 · p. 11
• Math-specific and tutoring models struggle to generalize to unseen or novel problems, while studies report downstream benefits from fine-tuning with preference data and instructions.Macina et al. (2023) found that small fine-tuned models outperform prompting a large model like ChatGPT in correctness and equitable tutoring.Source: §3.1.2 · p. 12

3.1.3 Instruction Following Tasks in Coding

• Coding evaluations model iterative generation, execution feedback, debugging, and refinement; newer benchmarks also target long-horizon interaction, security, and dynamic text-to-SQL.InterCode tests alternative interaction strategies, PyBench measures success and turns to completion, and coding research identifies unresolved choices about when to execute code, which protocols support refinement, and how to compare performance across languages and domains.Source: §3.1.3 · pp. 12–15

3.1.4 Discussions

• The discussion subsection draws on the benchmark tables and reviewed studies to identify trends in multi-turn instruction-following benchmarks and evaluation methods.Source: §3.1.4 · p. 15
• Benchmarks have expanded from MT-bench’s 80 examples to over 1,000 in newer datasets, while curation has shifted toward automated and LLM-assisted generation.The broader coverage addresses scalability while aiming to maintain quality.Source: §3.1.4 · p. 15
• Despite dataset growth, current benchmarks predominantly limit conversations to 10 or fewer turns, leaving dialogues lasting dozens or hundreds of turns largely unexplored.Source: §3.1.4 · p. 15
• Evaluation has shifted from rule-based metrics toward more nuanced, fine-grained frameworks, with LLM-as-a-judge emerging as a scalable, cost-effective approach.Source: §3.1.4 · p. 15
• MT-Bench (Zheng et al., 2023) pioneered LLM-as-a-judge evaluation while acknowledging its inherent biases.Source: §3.1.4 · p. 15
• LLM judges can show preference leakage toward responses from models sharing their lineage (Li et al., 2026b), inconsistent assessments of context-dependent outputs, and brittleness without references or when multiple solutions are valid.Source: §3.1.4 · pp. 15–16
• Human evaluation validation remains important, yet relatively few benchmarks substantially verify human–AI agreement, risking undetected biases and inconsistencies in automated evaluation.Source: §3.1.4 · p. 16

3.2 Conversational Engagement

• Conversational-engagement benchmarks span role-play, healthcare, education, and jailbreaks, with early work defining sustained dialogue quality through human labels and scalable LLM-based judging.ABC-Eval uses 16 binary turn-level behavior categories; BotChat uses LLM-generated conversations and judges, with GPT-4-generated dialogues nearly indistinguishable from human conversations.Source: §3.2 · p. 16
• Newer benchmarks test persistence and context management while auditing whether dialogue metrics and user simulators faithfully reflect human conversational performance.MultiChallenge covers instruction retention, implicit recall, iterative revision, and non-sycophantic consistency; DialogBench assesses naturalness and sustained interaction, and SimulatorArena tests simulator reliability.Source: §3.2 · p. 16

3.2.1 Conversational Engagement in Role-Play

• Role-play research uses persona prompts and behavioral presets to sustain character identity, and expert-role prompting can improve zero-shot accuracy across 12 tasks without parameter updates.CharacterChat combines an MBTI-based persona bank with context-specific memory; role-play prompting first generates an immersive persona acknowledgement and then answers using it.Source: §3.2.1 · pp. 17–18
• Role-play fine-tuning uses diverse persona-rich dialogue, character-profile-grounded data, and self-generated role-play conversations to improve in-character consistency and persona fidelity.The data ranges from PIPPA’s over 1 million messages to specialized character datasets; Ditto trains on dialogues it generates for 4,000 characters and frames role-play as reading comprehension to avoid style collapse.Source: §3.2.1 · p. 18
• Personalization and long-horizon role consistency use lightweight persona embeddings, context-gated character adapters, round-level state, offline reinforcement learning, and compressed conversational memory.PersonaPKT trains a persona vector using less than 0.1% of model parameters; MIDI-Tuning carries state across turns, while COMEDY summarizes history to help avoid contradictions and repetition.Source: §3.2.1 · p. 19
• Role-play evaluation covers character fidelity, temporal and social consistency, source-grounded character understanding, and personalization; specialized fine-tuned models can outperform GPT-4 on Chinese character consistency.CharacterEval evaluates 77 characters using 13 metrics; TimeChara finds timeline-boundary violations even among strong models, while RoleInteract reports that some agents struggle in group settings despite strong one-on-one performance.Source: §3.2.1 · pp. 19–20

3.2.2 Conversational Engagement in Healthcare

• Healthcare dialogue requires proactive multi-turn questioning to gather clinical details, retain conversation history, and integrate medical knowledge for accurate, context-appropriate responses.Studies use varied labels for this information-seeking capability, including chain of questions, proactive questioning, symptom inquiry, proactivity, and information seeking.Source: §3.2.2 · p. 23
• Healthcare systems combine dialogue-oriented medical training, term-aware generation, and patient-centered reasoning layers, while fine-tuned consultation models target medical and psychological dialogue.Clinical Camel uses dialogue-based knowledge encoding and outperforms GPT-3.5, GPT-4, and Med-PaLM2 on listed medical benchmarks; T-Agent improves ROUGE and term-extraction F1, while APP adds guidance without fine-tuning.Source: §3.2.2 · pp. 23–24
• BLEU-1 16.66 and ROUGE-1 27.44 on Huatuo-26M followed Qilin-Med’s DPO phase, improving beyond SFT; Aquila-Med used 12,727 preference pairs for medical consultations.Qilin-Med’s SFT improved accuracy over the Baichuan-7B baseline on CMExam.Source: §3.2.2 · p. 24
• AMIE uses simulated self-play and automated feedback, outperforming primary care physicians in diagnostic accuracy and earning higher ratings from specialists and patient actors.Its randomized, double-blind crossover evaluation used Objective Structured Clinical Examination scenarios.Source: §3.2.2 · p. 24
• Because conventional evaluation misses clinical complexity and safety, medical frameworks assess interactive quality, clinical reasoning, communication, and dynamic consultation behavior.MedGPTEval (Xu et al., 2023) assesses 27 multi-turn dialogues and 7 case reports; Liao et al. (2023) evaluates information gathering and diagnostic accuracy, while consultation-specific fine-tuning reduced hallucinations. AIE (Liao et al., 2024b), MMD-Eval (Liu et al., 2025d), and MediQ (Li et al., 2024b) use simulators or interactive benchmarks to assess dynamic dialogue and information seeking.Source: §3.2.2 · pp. 24–25
• MedFuzz (Ness et al., 2024) finds GPT-3.5, GPT-4, and Med-PaLM 2 perform worse on its stress-tested benchmark, exposing vulnerability to clinical biases, stereotypes, and incomplete data.Its adversarial inputs include ambiguous patient traits and biased data, challenging the assumption that high benchmark scores establish real-world reliability.Source: §3.2.2 · p. 25
• HealthBench (Arora et al., 2025) evaluates 5,000 real-world multi-turn healthcare conversations with more than 48,000 physician-developed criteria for accuracy, communication, uncertainty, and context awareness.It includes expert-validated Consensus and challenging Hard subsets.Source: §3.2.2 · p. 25
• Newer benchmarks test multi-turn degradation, premature answering, information lures, guideline adherence, long-horizon memory, and patient-facing consultation quality.MedDialogRubrics adds 5,200 synthetic cases and over 60,000 rubrics; MEDPI (V. et al., 2026) assesses safety and communication, while MindEval, Mazhar et al. (2026), MedAidDialog, and Dr. Assistant (Guo et al., 2026b) extend evaluation to mental health, multilingual consultations, and inquiry quality.Source: §3.2.2 · p. 25
• Healthcare models commonly use domain-specific medical dialogue data for SFT, sometimes followed by RL-based alignment, with evaluation extending beyond NLP metrics to clinical accuracy and human assessment.The survey attributes the pipeline’s effectiveness relative to broader dialogue settings partly to the quality and domain specificity of medical data.Source: §3.2.2 · p. 28

3.2.3 Conversational Engagement in Education

• Education-focused multi-turn systems fall into three areas: intelligent tutoring, automated grading and feedback, and scenario simulation.These systems use dialogue to guide students, iteratively assess and comment on their work, or simulate students and classrooms with AI agents.Source: §3.2.3 · p. 28

Socratic and Strategy-Guided Tutoring

• Socratic and strategy-guided tutors use reasoning prompts, targeted debugging questions, and planned teaching strategies to foster learner-led discovery.SocraticLM was trained on SocraTeach’s 35k multi-turn dialogues; TreeInstruct achieved state-of-the-art code-debugging results and students fixed bugs with minimal hints. StratL steered a Productive Failure tutor in a field experiment with 17 high school students.Source: Intelligent Tutoring Systems · pp. 28–29

Adaptive Tutoring

• PACE adapts Socratic explanations to modeled learning styles, while JeepyTA tracks course-forum questions and provides context-specific, conversational help across follow-up turns.PACE uses concrete examples for sensory learners and abstract prompts for intuitive learners; JeepyTA distinguishes logistical from conceptual questions and responds with hints or explanations.Source: Intelligent Tutoring Systems · p. 29

Evaluating LLM Tutoring in Mathematics

• MathDial evaluates faithful, equitable tutoring through simulated student misconceptions, teacher scaffolding, student success, and telling@k; fine-tuning on it improved student outcomes.SocraticMATH covers 513 primary-school math topics, while MathTutorBench assesses math expertise, student understanding, and teacher response quality.Source: Intelligent Tutoring Systems · p. 30
• Tutoring systems ground responses in course materials and optimize multi-turn replies for safety, pedagogical appropriateness, relevance, and student learning.CourseAssist uses syllabi and lecture notes; another approach ranks candidate replies with student and pedagogical evaluators, then fine-tunes by direct preference optimization.Source: Intelligent Tutoring Systems · p. 30

Bridging Research and Real Classrooms

• New benchmarks assess tutoring across K–8 math tasks, pedagogical dimensions, rubric-backed feedback, and safety failures that emerge in extended dialogues.KMP-Bench includes dialogue-based and verifiable tasks; MRBench audits 192 conversations and 1,596 responses across eight dimensions; TutorBench has 1,490 samples.Source: Intelligent Tutoring Systems · p. 30
• Classroom-facing datasets and educational models broaden deployment, but prompt moderation can underperform and realistic student simulation remains difficult.EduDial provides 34,250 sessions, TeachLM uses 100,000 hours of tutoring data, and ConvoLearn contributes 2,134 dialogues; LearnLM and Claude for Education illustrate industry adoption of specialized education models.Source: Intelligent Tutoring Systems · p. 30

Automated Feedback & Grading Support

• LLM feedback can sound insightful while being ungrounded, prompting methods that verify student reasoning, optimize for pedagogical preferences, or iteratively simulate revisions to improve drafts.Daheim et al. (2024) use step-by-step error analysis to guide targeted hints; Scarlatos et al. (2024) use rubric-based reinforcement learning; Nair et al. (2024) refine feedback through simulated student revisions.Source: Automated Feedback & Grading Support · pp. 30–31
• Educational deployments show both the value and limits of feedback: students found project-report comments helpful, but instructors wanted oversight, while database and medical tutors provided context-specific or supplemental feedback.Lohr et al. (2025) prompt feedback types from educational taxonomies; Dr.Academy found GPT-4 capable of generating higher-order questions aligned with Bloom’s taxonomy.Source: Automated Feedback & Grading Support · p. 31
• Over 95% accuracy on three-tier scoring and 98% on right/wrong judgments were reported for Spanish responses, while ChatGPT scores were within 10% for about 70% of exam responses.For the same 463 Master’s-level responses, 31% of ChatGPT scores were within 5%; it tended toward conservative scores, and GPT-4 struggled with nuanced essay rubrics and feedback.Source: Automated Feedback & Grading Support · p. 31
• Simulated learners span knowledge profiles, abilities, and personalities, enabling item analysis and multi-turn tests of tutors’ adaptation; SimClass extends this to classroom dialogue and peer interaction.TeachTune’s simulated-student behavior deviated from its profile by under 5–10%; SimClass user trials reported human-like exchanges and improved participant learning experience.Source: Automated Feedback & Grading Support · p. 32
• Scenario simulation also supports at-risk learner interventions, lesson-plan refinement, and synthetic chatbot training; Book2Dial’s LLM-based Role-Playing approach performed best among its dialogue-generation methods.Simulated interactions offer a sandbox for prototyping and testing educational approaches before involving real students.Source: Automated Feedback & Grading Support · pp. 32, 34

3.2.4 Conversational Engagement in Jailbreak

• Multi-turn jailbreaks exploit dialogue history through adaptive persuasion, escalating innocuous requests, or decomposing harmful prompts; PAIR often finds black-box attacks in fewer than twenty queries.Johnny / PAP reported attack success rates above 92% on Llama2-7B-Chat, GPT-3.5, and GPT-4; Crescendo relies on request order and implicit suggestions, while other approaches decompose harmful intent across turns.Source: §3.2.4 · pp. 34–35
• Jailbreak evaluation has expanded from adapting single-turn resources such as AdvBench and HarmBench to native multi-turn benchmarks, including SafeDialBench and human-authored MHJ.SafeDialBench contains more than 4,000 multi-turn safety dialogues across 22 scenarios and seven strategies; MHJ includes 2,912 prompts across 537 conversations with tactic labels and reviewer checks.Source: §3.2.4 · pp. 35–36
• Expert multi-turn jailbreaks exceed 70% attack success on HarmBench against defenses with single-digit rates under automated single-turn attacks, exposing a security–usability trade-off.Conventional filters and other defenses have limited efficacy against extended attacks; stronger defenses can reduce helpfulness or flag benign prompts.Source: §3.2.4 · p. 36

4 Improvements

• The improvement taxonomy has three branches—model-centric, external integration, and agent-based—and classifies overlapping works by their primary technical contribution.The branches respectively adapt the model, add resources such as memory or retrieval, and use systems that act, reflect, or collaborate; the taxonomy is an organizing lens, not a mutually exclusive partition.Source: §4 · pp. 37–38, Figure 3 · p. 38

4.1 Model-Centric Approaches

• Model-centric approaches address sustained interaction through in-context learning, supervised fine-tuning, reinforcement learning, and architectures designed for memory and dialogue flow.The section frames these methods as responses to the limits of techniques developed for static, single-turn queries.Source: §4.1 · p. 38

4.1.1 In-Context Learning

• Multi-turn prompting has context-dependent effects: naive demonstrations can hurt sequential reasoning, while task-aligned prompts and explicit dialogue structure can improve performance.A single demonstration reduced accuracy for some models on AQA-Bench coin-guessing, with recovery as examples increased; GraphIF uses relation graphs and Liu (2025) uses state updates to surface relevant context without training.Source: §4.1.1 · p. 39

4.1.2 Supervised Fine-Tuning

• Multi-turn SFT benefits from domain-specific dialogue data, while treating rounds independently risks context forgetting and incoherent responses.Reported gains span instruction following, role-play, and clinical dialogue, motivating fine-tuning methods tailored to multi-turn settings.Source: §4.1.2 · pp. 39–40
• SFT advances pair more realistic dialogue curation with training changes to loss design, context length, and supervision quality.Methods include simulator-generated or skeleton-guided dialogue, turn-wide loss objectives, extended contexts, and regularization of unreliable supervision signals.Source: §4.1.2 · p. 40

4.1.3 Reinforcement Learning

• RL results are mixed: specialized medical and coding methods show gains, but broad RLHF analyses find limited or negative multi-turn effects.LMRL-Gym and SWEET-RL provide evaluations for long-horizon interaction and collaboration, while newer approaches target turn-level proactivity and multi-turn jailbreak safety.Source: §4.1.3 · pp. 41–42
• A 7B model’s GSM8K accuracy rose from 77.5% to 83.9% after multi-turn preference optimization.M-DPO and M-KTO optimize tool-assisted reasoning trajectories; DMPO reformulates trajectory preference optimization, while ArCHer uses hierarchical credit assignment and is reported to outperform non-hierarchical baselines by approximately 100× in training samples.Source: §4.1.3 · pp. 41–42

4.1.4 New Architectures

• Memory and recurrence architectures extend context across turns, supporting dialogue coherence while reducing reliance on long input windows or fixed-length contexts.Examples include Cached Transformers’ gated recurrent cache, MemBART’s separately updated memory state, and recurrent designs that propagate information across segments.Source: §4.1.4 · pp. 42–43

4.2 External Integration Approaches

• External integration augments LLMs with memory, retrieval, or knowledge graphs to reduce compounding errors and misinformation, improving reliability, accuracy, and consistency in extended interactions.Source: §4.2 · p. 43

4.2.1 Memory-Augmented Methods

• Memory methods store and retrieve prior interactions to correct misunderstandings and support coherent, context-aware dialogue across turns.MemPrompt (Madaan et al., 2022) retrieves stored user corrections; Wu et al. (2025b) optimize memory indexing, retrieval, and reading, while later work explores episodic, hierarchical, and personalized memory designs.Source: §4.2.1 · pp. 43–44

4.2.2 Retrieval-Augmented Generation

• Multi-turn RAG grounds responses in retrieved documents and dialogue history, with recent systems modeling retrieval timing and which conversational states to retrieve.Benchmarks include MTRAG’s 110 conversations averaging 7.7 turns and 842 tasks, CORAL, and RAD-Bench’s 89 multi-turn question samples; RAGate (Wang et al., 2025b) also warns that indiscriminate retrieval can increase uncertainty and hallucination risk.Source: §4.2.2 · pp. 44–45

4.2.3 Knowledge Graph Integration

• Knowledge-graph methods add structured knowledge or dialogue state to LLM reasoning, supporting entity tracking, coreference, multi-hop reasoning, and grounded responses.Approaches include jointly trained graph and language models, Graph RAG retrieval of relevant subgraphs, and dialogue-conditioned structures such as D-SMART and GAP.Source: §4.2.3 · pp. 45–46

4.3 Agent-Based Approaches

• Agent-based approaches replace passive response generation with iterative reasoning, planning, and interaction, organized as single-agent or multi-agent systems.Source: §4.3 · p. 46

4.3.1 Single Agent Approaches

• Single-agent methods combine tools, action, and feedback: ReAct interleaves reasoning and action, while Reflexion and Voyager (Wang et al., 2024b) use iterative experience for improvement.Toolformer learns when to call tools, and HuggingGPT delegates subtasks; Voyager uses repeated planning, code execution, and observation to refine a skill repository.Source: §4.3.1 · pp. 46–47

4.3.2 Multi-Agents Approaches

• Role-based collaboration and dynamic team composition can improve performance: ChatDev reports significantly better software quality than single-agent baselines such as GPT-4, and AutoAgents surpasses fixed-role approaches on heterogeneous tasks.BUTTONInstruct contains 8,000 multi-turn dialogues with compositional function-calling tasks; models fine-tuned on it show significant improvements in planning and executing complex API call sequences.Source: §4.3.2 · pp. 47–48
• Multi-agent frameworks also risk role misassignment, communication overhead, propagated errors, weak verification, and conflicts or collusion, motivating stronger coordination and evaluation.Source: §4.3.2 · p. 48

5.1.1 Context Retention & Coherence

• Long conversations expose failures in retaining evolving context, resolving implicit references, and recognizing when ambiguity requires clarification.Models lose or conflate distributed state, struggle with cross-turn antecedents, and show an under-clarification bias in clarification benchmarks.Source: §5.1.1 · pp. 48–49, §5.1.2 · p. 49, §5.1.3 · p. 50

5.2.1 Error Propagation & Compounding

• Multi-turn reasoning remains brittle: early errors compound, topic shifts disrupt coherence, agents answer underspecified queries instead of asking, and code-switching undermines cross-lingual consistency.MINT finds that more than half of model answers are committed within the first half of a case, while accuracy improves when diagnosis is delayed until enough evidence accumulates; code-switching can also elicit undesirable behaviors.Source: §5.2.1 · p. 50, §5.2.2 · p. 50, §5.2.3 · p. 50

5.3.1 Dynamic Preference & Objective Adaptation

• Personalization and knowledge adaptation require stable yet revisable user representations and persistent learning, while staged attacks and misinformation make adaptation a safety risk.Current systems can forget or misapply preferences (s42.p1#9b679c), and true on-the-fly learning remains difficult amid catastrophic forgetting and malicious input risks (Wu et al., 2024a). Multi-turn safety can collapse under structured histories (Li et al., 2026c), and models may become more manipulable as they adapt.Source: §5.3.1 · p. 51, §5.3.2 · p. 51, §5.3.3 · pp. 51–52

5.4.1 Scalable Data Curation

• Scalable multi-turn data curation must improve dialogue quality and failure-mode coverage, not merely dataset size.Synthetic dialogues can lack spontaneity, amplify inaccuracies, and repeat generic patterns; the survey calls for better generation and filtering alongside shared multi-domain repositories (Qiu et al., 2024; Li et al., 2023a; Wang et al., 2024d; Yang et al., 2024; Ding et al., 2023; Maheshwary et al., 2025; Xu et al., 2024; Seddik et al., 2024; Chen et al., 2024f;c).Source: §5.4.1 · p. 52

5.4.2 Metric Design

• Evaluation needs turn-level and long-horizon measures of task success, fairness, and dialogue quality, while LLM judges require auditing for bias and multi-turn reliability.Single-turn accuracy and BLEU miss dialogue-level qualities; no universal fine-grained standard exists, and sustained performance measures remain underdeveloped. GPT-4 evaluators align with human preferences roughly 80% of the time but exhibit systematic biases (Zheng et al., 2023; Chen et al., 2024b); judge reliability can also drift across adversarial rounds.Source: §5.4.2 · pp. 52–53, §5.4.3 · pp. 53–54

5.5.1 Bias Amplification

• Multi-turn interaction can amplify demographic bias and enable iterative extraction of private or hidden information, leaving dialogue-level fairness and leakage prevention unresolved.LLMs accumulate bias more readily across turns than in single-turn prompts (Fan et al., 2025), while crafted prompt sequences can extract memorized data such as phone numbers and addresses.Source: §5.5.1 · p. 54, §5.5.2 · p. 54

5.5.3 Human–AI Bonding, Overtrust & Companion Risks

• Sustained personalization and emotional adaptation can foster dependence, overtrust, and unsafe validation, so safety must hold across interaction trajectories rather than isolated turns.A four-week randomized study links heavier chatbot use to greater emotional dependence and problematic AI use (Fang et al., 2025); companion dialogues can mirror affect, while counseling-like simulations may elicit toxic empathy and maladaptive validation. The survey identifies disclosure, boundary reminders, and escalation or handoff as potential mitigations (Akbulut et al., 2025; Boine, 2023).Source: §5.5.3 · p. 55

6 Conclusion

• The task-oriented taxonomy organizes interactions around instruction following and conversational engagement, covering domains such as role-play, healthcare, education, and jailbreak.It emphasizes coordinated capabilities in real-world tasks rather than isolated abilities such as reasoning, memory, or contextual understanding.Source: §6 · p. 55
• Multi-turn interaction more closely reflects sustained dialogue and iterative problem-solving than single-turn settings, requiring context retention, coherent cross-turn reasoning, adaptive behavior, and robust handling of evolving intent.The survey characterizes this as a fundamental shift in LLM utilization and evaluation.Source: §6 · pp. 55–56
• Enhancement approaches span model-centric methods, external integrations, and agent-based frameworks, with gains increasingly tied to managing interaction history and auxiliary state rather than model scale alone.Methods include in-context learning, fine-tuning, reinforcement learning, architectural innovations, memory augmentation, retrieval, knowledge graphs, and single- and multiagent systems.Source: §6 · p. 56
• Current challenges include extended-conversation coherence, cross-turn reasoning errors, topic-switching failures, limited adaptation, and uneven benchmark reporting across subdomains.The survey identifies dynamic preference learning and on-the-fly knowledge updates as particular adaptation limitations; dialogue scale, turn counts, evaluator setups, and agreement checks are reported unevenly.Source: §6 · p. 56
• Future progress requires stronger long-horizon context management, more reliable adaptation to users and tasks, and more auditable evaluation pipelines for high-stakes interactions.The survey’s taxonomy, domain analysis, methodological review, and organization of open challenges provide a foundation for this work.Source: §6 · p. 56

Broader Impact Statement

• Search and availability bias may underrepresent work from lower-resource regions, proprietary deployments, and communities using different terminology.The review is more likely to include work indexed in common scholarly databases, circulated on arXiv or major venues, written in English, and publicly accessible.Source: Broader Impact Statement · p. 56
• The survey’s non-multimodal focus, exclusion of MLLMs, and treatment of agentic systems as adjacent limit its coverage of multi-turn AI.These choices make the task-oriented synthesis more tractable, but the survey is not a complete account of all multi-turn AI systems; adjacent literatures are summarized in Appendix B.1.Source: Broader Impact Statement · p. 56
• The taxonomy’s primary-task or technical-contribution placements are interpretive choices, and alternative placements may also be defensible.Some papers mix benchmarks, methods, and analysis or span task and improvement categories, dialogue, agents, and multimodal interaction.Source: Broader Impact Statement · p. 56
• Emphasis on healthcare, education, roleplay, and jailbreak settings may understate other influential but less benchmarked or documented deployments.The authors frame the survey as a transparent, task-oriented synthesis rather than an exhaustive canon, and call for future work on multimodal, embodied, and more fully agentic systems, with inclusion criteria revisited as the field evolves.Source: Broader Impact Statement · p. 56

A Review Methodology Details

• The original corpus was assembled by April 2025 and extended through April 2026 using the same inclusion, exclusion, and scope rules.No prospective protocol was registered; the methodology is a retrospective transparency supplement, not a fully prospective systematic or scoping review.Source: A Review Methodology Details · p. 86
• Searches combined LLM, multi-turn, and task terms across venues, scholarly search tools, supplementary discovery, and revision updates, with section-specific queries and citation chasing.The latest update pass was in April 2026; searches included venue browsing, Google Scholar and ResearchGate, GPT-assisted discovery, and manual additions.Source: A Review Methodology Details · p. 86
• Included papers addressed non-multimodal multi-turn LLM interaction, contributed directly to the survey’s taxonomy or improvement and challenge sections, and provided sufficient technical detail.Potentially out-of-scope work was excluded only after receiving exclusion marks in three rounds; contextual non-paper resources were reported separately from papers in Figure 5.Source: A Review Methodology Details · pp. 86–87
• Manual screening recovered stage-level counts, used domain-oriented and additional reviewers plus multi-contributor section validation, and reported corpus counts by unique paper.The original exclusions were 188 single-turn-only, 64 environment-control or embodied, 52 agent-based, 27 multimodal, and 22 other scope mismatches; cross-cutting papers could appear in multiple sections.Source: A Review Methodology Details · p. 87
• For each included paper, authors manually charted bibliographic, task, role, benchmark, evaluation, findings, and limitation data, then checked table fields during drafting.Benchmark-related fields included dialogue counts, average turns, evaluator type, agreement checks, and evaluation criteria when reported; PRISMA-ScR was the primary reporting guide, with PRISMA 2020 supporting workflow transparency and limitations reporting (Page et al., 2021; Tricco et al., 2018).Source: A Review Methodology Details · pp. 87–88

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

• 275 items entered Version 1: 271 peer-reviewed or preprint papers and 4 contextual non-paper resources.The four resources were the AlpacaEval repository (Li et al., 2023e), LearnLM report, Claude for Education announcement, and USMLE sample-question resource (FSMB, 2023).Source: Figure 5 · p. 88, §A.1 · p. 88
• The workflow recorded 452 venue or database records, 216 scholarly-search records, and 37 supplementary additions; 648 reports were assessed for eligibility.Before eligibility assessment, 45 records were removed and 12 excluded at screening; 611 reports came from the database/register branch and 37 from supplementary discovery.Source: Figure 5 · p. 88, §A.1 · p. 88

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

• The reconstructed source-channel counts were 110 venue or database records, 9 from other scholarly search, 68 from two deep-research passes, and 35 manual additions.These subtotals are retrospective maintenance counts, not fixed prospective-review anchors.Source: §A.2 · p. 89
• Candidate papers were collected and screened under the same search, screening, and boundary-setting procedures, with the accumulated material consolidated after the TMLR reviews.The repository was actively maintained between versions rather than frozen as a prospective review log.Source: §A.2 · p. 89
• Figure 6 documents the revision workflow, which added 128 newly included studies to the 275 carried forward, bringing the Version 2 corpus to 403 studies overall.The retrospective update covered April 2025 through April 2026.Source: Figure 6 · p. 89, §A.2 · p. 89

B.1 Scope Boundaries and Adjacent Survey Directions

• The survey centers on non-multimodal multi-turn LLM interaction and distinguishes static dialogue evaluation, interactive tasks, and dynamic environment or agent settings.These boundaries are discussed alongside §2’s taxonomy.Source: §B.1 · p. 90
• Yi et al. (2025) survey recent LLM-based multi-turn dialogue systems, while Wang et al. (2023b) trace the shift from earlier dialogue systems to LLM-era settings.These surveys provide broader dialogue-system coverage than this survey aims to provide.Source: §B.1 · p. 90
• Healthcare reviews by Valizadeh & Parde (2022), Shi et al. (2024c), and Liu et al. (2024b) cover medical dialogue systems and medical LLM deployment more broadly.The passage identifies Liu et al. (2024b) as the Med-LLM survey.Source: §B.1 · p. 90
• Zhang et al. (2025a) is the closest existing overview, but frames multi-turn interaction by capabilities; Acikgoz et al. (2025) examine reasoning, monitoring, control, and long-horizon interaction.This survey uses a more explicitly task-oriented taxonomy and cites these works when they clarify capability-level distinctions.Source: §B.1 · p. 90
• Agentic LLM systems overlap through trajectories, memory, planning, and iterative feedback, but typically emphasize environment actions, tool use, or multi-agent coordination over plain-text dialogue.The survey discusses agentic work when relevant to multi-turn improvements but does not cover the agent literature exhaustively; broader examples include Xi et al. (2023), Guo et al. (2024), and Yehudai et al. (2025).Source: §B.1 · p. 90
• The survey excludes MLLMs because multimodal interaction brings different input spaces, task structures, and evaluation protocols, despite a rapidly growing benchmark ecosystem.Adjacent context includes ConvBench (Liu et al., 2024c), MMDU (Liu et al., 2024h), MMMT-IF (Epstein et al., 2024), MMDialog (Feng et al., 2023), TheaterGen (Cheng et al., 2024), and SVBench, which may inform future extensions.Source: §B.1 · p. 90

CE-Overview

• CE-Overview groups multi-turn dialogue evaluation benchmarks including ABC-Eval, BotChat, MultiChallenge, DialogBench, and SimulatorArena.Source: C Full Task Taxonomy · p. 92

CE-Roleplay

• CE-Roleplay covers persona modeling and role-play systems, consistency methods, and evaluation benchmarks, from PersonaChat to PERSONAMEM and OpenCharacter.Source: C Full Task Taxonomy · p. 92

CE-Healthcare

• CE-Healthcare spans domain-adapted and Chinese medical LLMs, medical knowledge sources, interactive diagnosis, multimodal and mental-health systems, consultation evaluation, and recent multi-turn benchmarks.Source: C Full Task Taxonomy · p. 92

CE-Education

• CE-Education catalogs pedagogical agents and math tutors alongside tutoring benchmarks, feedback and grading work, and simulated-student approaches.Source: C Full Task Taxonomy · p. 93

CE-Jailbreak

• CE-Jailbreak organizes foundational attacks, decomposition-based multi-turn exploits, safety benchmarks, recent multi-turn jailbreak research, defenses, and advanced attacks.Source: C Full Task Taxonomy · p. 93