Search and Filter

Beyond Single-Turn: A Survey On Multi-Turn Interactions With Large Language Models

Date

AI Overview

Why it matters

Real-world language model use increasingly requires sustained multi-turn exchanges rather than isolated single prompts. Evaluating multi-turn interactions helps address challenges such as maintaining context, handling dynamic goals, and preventing compounding errors across extended dialogues.

Study at a glance
  • Question: How are multi-turn interactions with large language models evaluated, improved, and challenged across various tasks?
  • Setting and data: 403 surveyed peer-reviewed and preprint studies published between 2022 and April 2026.
  • Method: Scoping literature review with a retrospective transparency protocol based on PRISMA-ScR.
  • AI tools: GPT-4, GPT-3.5, Claude, LearnLM, LLaMA, InstructGPT
Key findings
  • Most language models experience notable performance degradation over extended multi-turn exchanges because errors compound and context is lost across successive rounds.
  • Standard alignment techniques such as supervised fine-tuning and reinforcement learning from human feedback frequently produce uneven or limited improvements in multi-turn settings.
  • Automated model judges agree with human evaluators about 80% of the time, but they exhibit systematic position, verbosity, and self-enhancement biases.

Summary drafted with AI from the full paper and may contain errors. See the full paper for details.

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.