conversational ai

Existing research on large language model (LLM)-driven multi-turn dialogue systems suffers from fragmented frameworks, heterogeneous technical approaches, and inconsistent evaluation protocols. Method: This paper introduces, for the first time, a unified taxonomy of four LLM adaptation paradigms for multi-turn dialogue—prompt engineering, instruction tuning, retrieval-augmented generation (RAG), and dialogue state modeling—while distinguishing open-domain and task-oriented dialogue along technical and bottleneck dimensions. It constructs a structured technical landscape covering architectures, benchmark datasets, and evaluation metrics, and synthesizes multi-dimensional automatic and human evaluation methodologies. Contribution/Results: The survey identifies three critical research gaps—interpretability, long-horizon consistency, and controllable interaction—and establishes a theoretically grounded, practice-oriented reference framework to guide future work in LLM-based dialogue systems.

This study addresses the weak interactive capability of large language models (LLMs) in multi-turn dialogue. Methodologically, it introduces a unified “capability–evaluation–enhancement–evolution” analytical framework—the first to jointly characterize context retention, dynamic response generation, and interactive modeling. The approach integrates dialogue state tracking, long-context modeling, trajectory-driven evaluation metrics, retrieval-augmented generation, and memory mechanisms, yielding a scalable multi-turn evaluation protocol and a collaborative interactive evolution pathway. The work provides the first systematic survey of LLM interactivity, precisely identifying core bottlenecks—including state drift, long-range forgetting, and evaluation misalignment—and offers both theoretical foundations and practical guidelines for applications such as conversational search, intelligent consulting, and interactive pedagogy.

Large language models (LLMs) exhibit unstable dialogue behavior and poor maintainability in complex business processes. Method: This paper proposes Conversation Routines (CR), a framework that formalizes task-oriented dialogue logic via natural-language specifications, pioneering the integration of structured business workflows directly into LLM prompts—thereby decoupling dialogue design from tool implementation. CR supports modular routine definition and composition, natural-language-driven workflow orchestration, and synergistically combines tool-augmented conversational agents (Tool-Augmented CAS) with prompt engineering. Contribution/Results: Evaluated on two proof-of-concept scenarios—train ticket booking and interactive fault diagnosis—CR enables domain experts to build high-fidelity, high-task-success-rate dialogues without coding. It significantly improves system interpretability, reusability, and cross-role collaboration efficiency.

Goal-oriented open-domain dialogue systems struggle to simultaneously achieve user personalization, phase adaptability, and low-data learning. Method: This paper proposes a novel framework integrating large language models (LLMs) with a hierarchical reinforcement learning (HRL)-based dialogue manager. Specifically: (i) HRL models multi-phase dialogue policies to enable smooth, goal-driven transitions across phases; (ii) a meta-learning mechanism enables rapid personalization across diverse user profiles; and (iii) an LLM–HRL co-architecture decouples semantic generation from policy decision-making, reducing reliance on annotated dialogue data. Results: Evaluated on motivational interviewing tasks, the proposed dialogue manager achieves significantly higher reward scores than state-of-the-art LLM-based baselines, demonstrating superior performance in goal completion rate, user adaptability, and data efficiency.

This work addresses inherent limitations of large language models (LLMs) in semantic understanding and behavioral reliability. We propose a hierarchical dialogue architecture that tightly couples LLMs with Answer Set Programming (ASP), wherein the LLM serves *exclusively* as a bidirectional parser between natural language and formal logic, while all core logical inference is delegated to ASP. This strict separation ensures semantically interpretable and behaviorally verifiable dialogue reasoning. Our key contribution is the first principled decoupling framework—LLM+ASP—that explicitly disentangles linguistic interpretation from symbolic reasoning, thereby overcoming critical bottlenecks in end-to-end LLM-based dialogue systems: pervasive hallucination, ill-defined operational boundaries, and non-auditable reasoning traces. Empirical evaluation on both task-oriented and social dialogue prototypes demonstrates significant improvements in reasoning traceability, hallucination resistance, and behavioral controllability.

Domain-specific chatbots suffer from ambiguous user intent, contextual fragmentation, and interaction disorganization during multi-turn interactions—such as conditional filtering, multi-option selection, and comparative operations—due to the absence of GUI-like “submit/reset” mechanisms. To address this, this work introduces, for the first time, a form-based Submit/Reset paradigm into conversational systems, explicitly modeling user confirmation behaviors and context-switching actions. Methodologically, we integrate formalized state tracking, fine-grained user action recognition, and chain-of-thought (CoT) reasoning, augmented by prompt engineering to enhance large language models’ capacity for structured dialogue state representation. Experiments in hotel booking and customer management domains demonstrate significant improvements: +28.6% in multi-turn task coherence, +32.1% in user satisfaction, and a reduction of 2.4 turns on average, indicating enhanced operational efficiency.

To address the challenges of discontinuous intent understanding and context-agnostic responses in conversational search under complex information needs, this paper proposes an LLM-driven multi-turn conversational search framework. Methodologically, it synergistically integrates classical IR principles with native LLM capabilities—including instruction following and chain-of-thought reasoning—through dialogue state tracking, hierarchical prompt engineering, lightweight instruction fine-tuning, and an interpretable reasoning module, enabling context-aware intent modeling and dynamic response generation. Key contributions include: (1) the first systematic characterization of foundational paradigm shifts in conversational search in the LLM era; (2) the establishment of a comprehensive research framework spanning theoretical analysis, technical implementation, and empirical validation; and (3) significant improvements in multi-turn relevance and response consistency across multiple complex QA benchmarks, delivering a reusable, next-generation intelligent conversational search paradigm for both academia and industry.

This work addresses core challenges in transitioning speech dialogue systems from cascaded ASR/NLU pipelines to end-to-end multimodal architectures. Method: We propose a speech-native large language model (LLM) framework integrating audio-adapted LLMs, cross-modal alignment mechanisms, joint speech-text pretraining, streaming inference, post-ASR correction, and enhanced multilingual accent robustness. Contribution/Results: (1) A unified architectural perspective bridging industrial voice assistants and open-domain/task-oriented agents; (2) Release of reproducible baselines, a system-level development roadmap, and standardized evaluation protocols; (3) Comprehensive curation of key datasets and explicit identification of open challenges—including privacy, safety, and rigorous multimodal evaluation. Our framework advances spoken dialogue systems toward next-generation architectures that are more robust, scalable, and inherently multimodal.

To address the challenge of natural language interaction with industrial-scale ERP systems, this paper proposes a dual-agent collaborative architecture: a reasoning agent generates SQL queries, while a critique agent iteratively validates and refines them using database schema knowledge and execution feedback. The approach integrates open-source large language models, domain-adapted NL2SQL fine-tuning, and a dynamic execution-feedback mechanism, significantly enhancing SQL generation accuracy and robustness. Experiments in a real ERP production environment achieve 92.3% intent understanding accuracy and an 89.7% executable SQL rate—improving upon baseline models by 14.5 percentage points. Key contributions include: (1) the first verifiable dual-agent NL2SQL framework specifically designed for industrial ERP scenarios; (2) an execution-feedback-driven self-correction mechanism enabling iterative query refinement; and (3) empirical validation of the practical viability of lightweight open-source models in complex enterprise systems.

In task-oriented dialogue, user utterances are often semantically complete yet lack system-executable structured intents—a challenge arising from the asymmetry between users’ ambiguous needs and systems’ requirement for precise intent definitions. To address this, we propose STORM, the first framework to formally model the dynamic evolution of intent triggerability under such asymmetric interaction. STORM introduces an intent formation trajectory model to characterize the cognitive progression of collaborative understanding and designs a novel evaluation metric jointly optimizing cognitive grounding and task performance. Methodologically, it employs a dual-LLM architecture (UserLLM/AgentLLM), structured trajectory annotation, and uncertainty-sensitive evaluation. Evaluated on four mainstream LLMs, STORM demonstrates that strategies incorporating 40–60% moderate uncertainty significantly outperform fully transparent ones, revealing model-specific uncertainty calibration patterns. These results provide both theoretical foundations and empirical evidence for enhancing cooperative dialogue systems through calibrated uncertainty modeling.

Intent classification and out-of-scope (OOS) detection in task-oriented dialogue systems heavily rely on large-scale annotated data, posing a critical bottleneck in low-resource settings. Method: This paper proposes a hybrid framework synergizing BERT and large language models (LLMs). Under zero-shot and few-shot settings, BERT efficiently extracts semantic features, which are structured and injected as prompt-enhanced inputs to the LLM, enabling cross-model information fusion and reasoning optimization. Contribution/Results: A lightweight feature-sharing mechanism is introduced—improving LLM generalization without increasing its parameter count. Experiments on multi-turn, multi-party dialogue datasets demonstrate that, using only 5% labeled data, our method achieves absolute accuracy gains of +12.3% for intent classification and +15.7% for OOS detection over strong baselines. The approach effectively alleviates annotation scarcity and establishes a new paradigm for robust dialogue understanding under data-constrained conditions.

Contemporary dialogue systems typically adopt an integrated architecture combining large language models (LLMs), external tools, and databases; thus, evaluating only the underlying LLM fails to ensure end-to-end quality. Existing evaluation methods predominantly focus on single-turn analysis and lack automated, process-aware testing for full conversational trajectories. Method: We propose the first end-to-end testing framework based on non-cooperative user simulation: (1) a challenging, role-driven user simulator requiring no reference dialogues or system-internal knowledge; (2) a fine-grained error taxonomy to guide prompt optimization and enhance detection of dialogue failures and anomalies; and (3) a decoupled architecture enabling low-cost configuration and cross-system portability. Contribution/Results: Experiments demonstrate substantial improvements in defect detection rates, with strong generalizability, scalability, and robustness across diverse dialogue systems and evaluation settings.