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

1. Problem Statement

While Large Language Models (LLMs) have achieved remarkable success in single-turn tasks (isolated prompts), real-world applications increasingly demand multi-turn interactions. Single-turn evaluations fail to capture critical complexities inherent in sustained dialogues, such as:

• Context Retention: The ability to remember and utilize information from previous turns over long sequences.
• Coherence & Consistency: Maintaining logical flow, persona consistency, and avoiding contradictions across turns.
• Dynamic Adaptation: Handling evolving user intents, ambiguous inputs, and iterative goal refinement.
• Error Propagation: The tendency for early errors to compound and derail subsequent reasoning.
• Safety & Fairness: Increased vulnerability to adversarial jailbreaking (e.g., gradual manipulation) and bias amplification over time.

Current literature lacks a holistic, task-oriented synthesis of how to evaluate, improve, and deploy LLMs specifically for these prolonged, interactive scenarios.

2. Methodology

The authors conducted a comprehensive survey of recent literature, organizing the field through a task-oriented taxonomy rather than a capability-oriented one. This approach reflects real-world deployment where multiple capabilities (reasoning, memory, planning) must intertwine to achieve specific goals.

The survey is structured into three main pillars:

1. Task Categorization: Multi-turn interactions are divided into Instruction Following (clear user intent, e.g., math, coding) and Conversational Engagement (open-ended, e.g., role-play, healthcare, education, jailbreaking).
2. Improvement Methodologies: A systematic review of techniques to enhance multi-turn performance, categorized into:
   • Model-Centric: In-context learning, Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and new architectures.
   • External Integration: Memory augmentation, Retrieval-Augmented Generation (RAG), and Knowledge Graphs.
   • Agent-Based: Single-agent and multi-agent collaborative systems.
3. Challenge Analysis: Identification of open research problems across six dimensions: Context Understanding, Complex Reasoning, Adaptation & Learning, Evaluations, and Ethical & Safety Issues.

3. Key Contributions

A. Task-Oriented Taxonomy

The paper introduces a novel framework categorizing multi-turn tasks based on user intention clarity and task complexity:

• Instruction Following Tasks:
  • General: Mixed domains (MT-Bench, MT-Bench++).
  • Specialized: Mathematics (MathChat-Bench, MINT) and Coding (InterCode, MMSQL).
  • Emerging: Fairness (FairMT-Bench) and Feedback responsiveness (FB-Bench).
• Conversational Engagement Tasks:
  • Role-Playing: Focus on persona consistency, memory, and emotional depth (e.g., CharacterEval, RoleEval).
  • Healthcare: Proactive symptom inquiry, diagnostic reasoning, and empathy (e.g., AMIE, HuaTuoGPT).
  • Education: Socratic tutoring, adaptive feedback, and scenario simulation (e.g., MathDial, TeachTune).
  • Jailbreaking: Multi-turn adversarial attacks (e.g., Crescendo, ActorAttack) and defenses.

B. Comprehensive Review of Improvement Strategies

The survey details three distinct methodological branches:

1. Model-Centric Approaches:
   • In-Context Learning (ICL): Analysis of how multi-turn exemplars can sometimes degrade performance if not aligned with task dynamics.
   • Supervised Fine-Tuning (SFT): Emphasis on curating realistic multi-turn datasets (e.g., self-chat, synthetic dialogues) and optimizing loss functions for sequence coherence.
   • Reinforcement Learning (RL): Shift from single-turn RLHF to Multi-Turn RL (e.g., DMPO, M-DPO, ArCHer) that optimizes entire conversation trajectories and handles credit assignment across turns.
   • New Architectures: Exploration of cached transformers, memory-augmented models (MemBART), and recurrent structures (RWKV) to handle long-range dependencies.
2. External Integration Approaches:
   • Memory Augmentation: Techniques like MemPrompt and MemTree to store and retrieve episodic dialogue history.
   • RAG & Knowledge Graphs: Using external retrieval and graph structures (e.g., Graph RAG) to ground responses in factual knowledge and resolve entity references.
3. Agent-Based Approaches:
   • Single Agents: Frameworks like ReAct and Reflexion that enable iterative planning, tool use, and self-correction.
   • Multi-Agent Systems: Collaborative frameworks (e.g., CAMEL, ChatDev, MetaGPT) where specialized agents debate or divide tasks to solve complex problems.

C. Identification of Open Challenges

The paper categorizes critical gaps into six areas:

• Context Understanding: Issues with anaphora resolution, ellipsis, and maintaining coherence over long contexts.
• Complex Reasoning: Error propagation, topic switching, and the lack of proactive information seeking.
• Adaptation & Learning: Difficulty in dynamic preference adaptation and on-the-fly knowledge updates without catastrophic forgetting.
• Evaluations: The need for scalable data curation, better metrics beyond single-turn accuracy, and addressing the bias in LLM-as-a-judge systems.
• Ethical & Safety: Risks of bias amplification, privacy leakage through iterative probing, and the "hyper-realism" problem (users over-trusting AI).

4. Results & Findings

• Performance Degradation: State-of-the-art models often show significant performance drops in multi-turn settings compared to single-turn, particularly in context retention and consistency.
• Training Necessity: Standard single-turn alignment (SFT + RLHF) is insufficient for multi-turn robustness. Specialized multi-turn training (e.g., trajectory-level RL, multi-turn SFT) yields superior results.
• Evaluation Gaps: Current benchmarks are often limited to short dialogues (<10 turns) and rely heavily on automated judges (LLM-as-a-judge), which suffer from biases (e.g., preference leakage, verbosity bias).
• Domain Specificity: In high-stakes domains like healthcare and education, multi-turn capabilities (proactive questioning, Socratic guidance) are essential and show significant gains when models are fine-tuned on domain-specific dialogue data.
• Security Risks: Multi-turn interactions create new attack vectors (e.g., "Crescendo" attacks) where harmless initial prompts are gradually manipulated to bypass safety filters, a vulnerability not present in single-turn settings.

5. Significance

This survey is significant because it is the first comprehensive work to bridge the gap between single-turn LLM capabilities and the complex realities of multi-turn deployment.

• Paradigm Shift: It moves the research focus from isolated capability testing to task-oriented interaction, acknowledging that real-world utility depends on the interplay of memory, reasoning, and adaptation.
• Roadmap for Researchers: By systematically categorizing benchmarks, methods, and open challenges, it provides a clear roadmap for future research in making LLMs more robust, safe, and effective in sustained conversations.
• Practical Impact: The insights are crucial for deploying LLMs in critical sectors (healthcare, education, customer service) where long-term context and reliability are non-negotiable.
• Resource Availability: The authors provide a curated repository of resources and papers, facilitating further research and development in this rapidly evolving field.

In conclusion, the paper argues that mastering multi-turn interactions is the next critical frontier for LLM evolution, requiring a holistic approach that integrates advanced modeling techniques, external knowledge, and rigorous, domain-specific evaluation frameworks.