Beyond Dyadic Dialogue: A Comprehensive Survey of Multi-Party Dialogue Systems

1. Introduction

Dialogue systems form a foundational pillar of natural language processing (NLP). From early rule-based assistants to modern LLM-powered conversational agents, the field has primarily targeted dyadic interactions: a single user engaging with a single system. Typical applications include customer service, question-answering bots, task-oriented dialogue, and knowledge-grounded healthcare dialogue such as cognitive stimulation systems for elderly users Jiang et al. (2023). With a limited number of participants and relatively straightforward turn structure, these systems have benefited from modeling paradigms for tracking conversational state, contextual grounding, and response generation Vaswani et al. (2017).

However, dyadic settings capture only a slice of human communication. In meetings, classrooms, online forums, and collaborative platforms, conversations typically involve multiple participants who exchange information, negotiate intent, and share knowledge in parallel. Such Multi-Party Dialogue (MPD) is fundamentally more complex Ishizaki and Kato (1998); Sapkota et al. (2025). Three dimensions distinguish MPD from its dyadic counterpart. First, the participant dimension requires speaker tracking, addressee recognition, and turn management. Second, the discourse structure dimension involves frequent topic shifts, interleaved threads, and multi-level discourse organization. Third, the task dimension demands modeling of social roles, intent dynamics, and recurring interaction patterns alongside coherent generation.

Across these dimensions, MPD research has progressed through several methodological waves. Early systems relied on statistical and rule-based methods such as Hidden Markov Models, Conditional Random Fields, and finite-state machines for dialogue state tracking and intent recognition Martínez-Hinarejos et al. (2010); Shang et al. (2020); Zhu et al. (2010). The deep learning era introduced CNN, RNN, and LSTM architectures that improved nonlinear context modeling Mangrulkar et al. (2018); Wen et al. (2015); Skantze (2017). More recently, large language models such as GPT, T5, LLaMA, and Qwen have shown strong language understanding and generation capabilities, providing new affordances for long-context, multi-speaker reasoning. The latest frontier extends LLMs into multi-agent and hybrid paradigms, enabling autonomous role assignment, information sharing, and human-machine collaboration in group settings Sapkota et al. (2025); Wang et al. (2026).

Despite this progress, the field faces persistent obstacles. Existing models struggle to manage combinatorial information flow among speakers and to maintain coherent discourse structure across interleaved threads. Datasets remain limited in scale, modality, and annotation depth compared to dyadic resources. Evaluation protocols largely inherit metrics such as BLEU and F1 from NLP tasks, which fail to capture the social, structural, and collaborative qualities that matter in group conversation. Together, these gaps slow both methodological maturation and practical deployment.

This survey synthesizes the current state of MPD research with the goal of clarifying where the field stands and where it should go. Our contributions are as follows: (1) We propose a unified taxonomy of MPD tasks spanning six categories, from participant modeling to high-level decision support (Section 2). (2) We chart the methodological trajectory of MPD systems, organizing approaches from statistical models, through neural and graph-based methods, to modern LLM and multi-agent paradigms, with a complete catalog of 50+ methods (Section 3). (3) We catalogue representative datasets and evaluation metrics, highlighting their coverage, limitations, and complementarity (Section 4 and Section 5). (4) We articulate four fundamental challenges facing MPD research and map each to concrete opportunities, providing a strategic roadmap for the next generation of socially intelligent dialogue systems (Section 6).

2. Tasks in Multi-Party Dialogue

MPD encompasses a diverse landscape of tasks that reflect its inherent multi-speaker, multi-thread complexity. Figure 1(a) provides a visual taxonomy. We organize existing tasks into six interdependent groups, presenting task definitions and representative work for each.

2.1. Dialogue Understanding

Dialogue understanding tasks aim to capture the semantic content of multi-party dialogues, such as goals, emotions, and speaker intentions. These tasks provide the foundation for higher-level modeling and reasoning.

Goal Tracking and Intent-Slot Recognition. This task involves detecting and tracking conversational goals as well as identifying semantic slots. In multi-party contexts, the difficulty arises from overlapping goals among participants and potential goal shifts during interactions Addlesee et al. (2023); Penzo et al. (2024). Recent work explores hybrid intent recognition frameworks that combine lightweight models such as BERT with LLMs under zero- and few-shot settings. By leveraging uncertainty-based routing and sharing probability-based information between models, these approaches improve intent recognition and out-of-scope detection performance in multi-party conversations while reducing computational costs Castillo-López et al. (2025).

Emotion Recognition in MPD. Recognizing emotions across multiple participants is crucial for building socially intelligent systems. Emotion dynamics in MPD are more complex than in dyadic dialogues, as emotional contagion and group-level sentiment often emerge Sun et al. (2021).

Dialogue Act Classification. Assigning functional roles (for example, question, answer, suggestion, agreement) to utterances supports downstream applications such as summarization and decision support. In MPD, dialogue act recognition must consider both local context and cross-speaker dependencies Martinenghi et al. (2024); Mayfield et al. (2012).

Multi-party Dialogue Reading Comprehension (MRC). MPD-MRC extends traditional machine reading comprehension to dialogue transcripts. It requires reasoning over multiple turns and speakers, integrating discourse and role information for accurate question answering Li and Zhao (2023, 2021).

Disruptive Talk Detection. Detecting disruptive or toxic utterances is important for maintaining constructive discussions in collaborative or educational settings Park et al. (2022).

Argument Prediction in Debates. In debate-style MPD, predicting which arguments are more persuasive or likely to win is a unique task with implications in computational social science Sia et al. (2022).

2.2. Dialogue Structure Modeling

These tasks aim to capture and represent the structural properties of MPD, which are often more complex than in dyadic dialogues.

Turn-taking Prediction and Modeling. MPD requires predicting when a participant will speak and determining the next speaker. This is challenging due to overlapping speech, interruptions, and implicit floor management strategies Gatti de Bayser et al. (2019); Laskowski (2010); Enomoto et al. (2020); Wang et al. (2024).

Topic Segmentation and Identification. Identifying topic boundaries and recognizing emerging topics in multi-party discussions is crucial for summarization and retrieval. Unlike single-topic dialogues, MPD often involves multiple threads of conversation Purver et al. (2006); Galley et al. (2003); Nguyen et al. (2013).

Dynamic Topic Disentanglement. Conversations in MPD are often interleaved. Disentanglement aims to separate intertwined threads into coherent topic-specific sub-dialogues Wang et al. (2020).

Discourse Parsing and Structure Prediction. Parsing discourse relations (for example, elaboration, contrast, agreement) and predicting conversation structures are essential for understanding the logical flow of MPD. Successful entry prediction, segmentation, and coherence modeling fall under this area Afantenos et al. (2015); Li et al. (2023); Liu and Chen (2021); Liu et al. (2025); Mayfield et al. (2012).

Decision-making Sub-dialogue Detection. Many MPD scenarios, such as meetings, involve decision-making. This task detects sub-dialogues where group decisions are negotiated Fernández et al. (2008); Bui et al. (2009).

Multi-turn Alignment and Chain-of-Thought Reasoning. Aligning utterances across turns and modeling explicit reasoning chains in MPD enable advanced reasoning-oriented applications Kiruluta et al. (2025).

2.3. Response and Generation

Response and generation tasks are at the heart of dialogue system development. In MPD, response generation is complicated by multiple potential addressees, shifting contexts, and social constraints.

Response Generation in MPC. Generating contextually appropriate responses in multi-party dialogues is one of the most fundamental challenges. It requires reasoning about speaker history, dialogue context, and social dynamics Fan et al. (2024); Gu et al. (2022); Gu et al. (2023); Hu et al. (2025); Li et al. (2023); Liu et al. (2019); Tan et al. (2023); Wang et al. (2020).

Response Selection. Given multiple candidate responses, the task is to select the most appropriate one. This often serves as an intermediate step for evaluating dialogue models Gu et al. (2021a) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2023); Ouchi and Tsuboi (2016); Penzo et al. (2024); Wang et al. (2020).

Response Generation with Addressee Prediction. Unlike dyadic dialogues, MPD responses must often include an explicit addressee, requiring models to jointly predict the response and its recipient Li and Zhao (2023).

Persona-aware Generation. This task focuses on maintaining speaker-specific personalities or roles across responses, which is especially relevant in long-term or role-based conversations Ju et al. (2022); Mahajan and Shaikh (2024).

Empathetic Dialogue Generation in MPD. Empathy is particularly complex in MPD, as systems must balance the emotions of multiple participants while maintaining coherence Zhu et al. (2022). This challenge is exemplified by group cognitive stimulation dialogue for elderly users with cognitive impairment, where an adaptive policy must coordinate therapeutic goals across several participants simultaneously Jiang et al. (2026).

Dialogue Generation with Next-Speaker Prediction. Extends response generation to also predict which participant should speak next, aligning generation with realistic conversation dynamics Wang et al. (2024).

Tool-calling in Multi-party Dialogue. Emerging work incorporates external APIs or tools (for example, calendars, databases) during multi-turn group conversations, making dialogue more interactive and actionable Jang et al. (2025). Beyond static tool invocation, recent work explores tool-driven intrinsic adaptation, where agents unify task execution with self-reconfiguration of their tool repertoire Zhou et al. (2026).

2.4. Role and Participant Modeling

Since MPD inherently involves multiple participants, modeling their roles, identities, and interactions is a crucial research focus.

Addressee Recognition and Identification. Identifying the intended recipient of an utterance is a fundamental task in MPD. It is closely tied to response generation and turn-taking modeling Gu et al. (2021a) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2023); Heylen and op den Akker (2007); Le et al. (2019); Ouchi and Tsuboi (2016); Penzo et al. (2024); Zhu et al. (2023).

Speaker Identification and Classification. Determining who is speaking in overlapping or noisy conditions is a prerequisite for many higher-level tasks Du et al. (2022); Gu et al. (2021a) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2023); Meng et al. (2018).

Overlapped Speaker Diarization. This task aims to separate overlapping speech and attribute it correctly to different participants, which is especially challenging in real-time settings Du et al. (2022).

Social Role and Seniority Inference. Inferring roles such as leader, moderator, or newcomer provides insights into group dynamics and social hierarchy Laskowski et al. (2008).

Interaction Pattern Analysis. Beyond individual roles, this task models recurring interaction structures, such as supportive behaviors, challenges, or questioning patterns Heylen and op den Akker (2007).

2.5. Summarization and Decision Support

Summarization and decision-support tasks seek to condense lengthy dialogues into concise, actionable outputs. These are particularly relevant in professional or collaborative contexts.

Dialogue Summarization. Generating abstractive or extractive summaries of group discussions. This requires modeling global context, topic transitions, and participant contributions Chen and Yang (2020).

Extractive Summarization in MPC. A variant of summarization that selects salient utterances instead of generating new text Chen and Metze (2012).

Automatic Minuting. Specialized summarization that produces structured meeting minutes with decisions, action items, and responsibilities Bhatnagar et al. (2022).

Decision Discussion Detection and Summarization. Identifying decision-related dialogues and producing concise summaries tailored to decision-making processes Bui et al. (2009).

2.6. Advanced and Emerging Tasks

Beyond classical tasks, researchers have explored advanced and domain-specific problems in MPD.

Machine Translation in MPD. Translating utterances in multilingual multi-party conversations while preserving role and context Bhatnagar et al. (2022).

Extractive Question Answering in MPC. Answering factual questions directly from conversation transcripts Li et al. (2023); Zhou et al. (2025).

Negotiation Strategy Learning. Learning strategies for successful negotiation in multi-party discussions, applicable to economics and multi-agent systems Hiraoka et al. (2015).

Entity Linking in MPC. Linking entity mentions within dialogues to external knowledge bases, facilitating knowledge-grounded conversations Aina et al. (2019).

Crucially, these tasks are deeply interdependent. Accurate addressee recognition strengthens response generation, while reliable discourse parsing improves summarization. This interdependence motivates the recent trend toward holistic, multi-task modeling discussed in Section 6.

3. Methods in Multi-Party Dialogue

The methodological history of MPD mirrors the broader arc of NLP, progressing from statistical models, through neural and graph-based architectures, to large language models and multi-agent systems. Figure 2 visualizes this evolution along four major research branches that have continuously matured over two decades, while Table 1 highlights representative methodological milestones along the trajectory. The complete catalog of 50+ methods appears in Table 4, grouped chronologically into the statistical/early-neural era (pre-2020), the pre-trained model and graph era (2020–2023), and the LLM and multi-agent era (2024–2026).

3.1. From Statistical Models to Neural Networks

Early MPD systems leveraged statistical learning. Classical models for next-speaker prediction (MLE, SVM, CNN) demonstrated the value of semantic features in turn dynamics Gatti de Bayser et al. (2019). Beyond turn dynamics, statistical approaches addressed a wide range of structural and social tasks. Unsupervised Bayesian generative topic models offered robustness to ASR noise and domain shift for topic segmentation in noisy meeting transcripts Purver et al. (2006), while feature-driven segmentation combined lexical and acoustic cues for domain-agnostic boundary detection Galley et al. (2003). ILP formulations with sociolinguistic constraints tackled hierarchical structure prediction Mayfield et al. (2012), MST-based dependency parsing modeled discourse structure in multi-party chat Afantenos et al. (2015), and structured labeling together with directed-graph methods detected and summarized decision-related sub-dialogues Fernández et al. (2008); Bui et al. (2009). Complementary lines of work predicted social dimensions and seniority from interaction patterns Laskowski et al. (2008), analyzed backchannel and gaze cues for interaction modeling Heylen and op den Akker (2007), applied random-walk extractive summarization to meeting transcripts Chen and Metze (2012), learned negotiation strategies in simulated trading via reinforcement learning Hiraoka et al. (2015), and linked entity mentions to knowledge bases Aina et al. (2019).

The shift to neural architectures introduced richer context modeling. Joint static-dynamic RNN models tackled addressee and response selection together Ouchi and Tsuboi (2016), W2W learned to identify utterance addressees from raw context Le et al. (2019), and neural speaker-classification models compared temporal, content-driven, and hybrid representations, revealing the benefit of integrating sequential and semantic information Meng et al. (2018). Context encodings that integrate speaker and addressee role information improved response generation Liu et al. (2019), and BERT-based multi-task models enabled dynamic topic tracking for reply selection Wang et al. (2020). These neural advances laid the groundwork for the pre-trained model era that followed.

3.2. Pre-trained Models and Graph Reasoning

Pre-trained language models reshaped MPD research. MPC-BERT jointly models speaker identity and utterance semantics through self-supervised tasks including reply-to recognition, identical speaker searching, and pointer consistency distinction Gu et al. (2021a) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang, while latent discourse inference exploits unlabeled multi-party data efficiently Li et al. (2023). Pseudo-self-supervised pre-training likewise improves multi-party reading comprehension and dialogue QA Li and Zhao (2021), and cross-domain Transformer parsers transfer discourse-parsing capability across corpora Liu and Chen (2021). To address the graph-structured nature of MPD, heterogeneous graph neural networks such as HeterMPC Gu et al. (2022) and MADNet Gu et al. (2023) explicitly model utterances and interlocutors as typed nodes with multiple meta-relations, with EM-based optimization handling missing addressee data. Lightweight graph-induced fine-tuning (GIFT) integrates structural signals into pre-trained Transformers with only a handful of additional parameters per encoding layer Gu et al. (2023), and persona-aware variants enhance speaker consistency through personality-based embeddings Mahajan and Shaikh (2024); Ju et al. (2022). Global-local graph reasoning further supports multi-hop question answering over complex multi-party contexts Li and Zhao (2023), and graph-aware sentiment models capture emotion propagation across speakers Sun et al. (2021). The same period also saw multi-view BART for abstractive summarization Chen and Yang (2020), static-dynamic empathy frameworks (SDMPED) for emotion-aware generation Zhu et al. (2022), discrete addressee codebook encodings for robust addressee recognition Zhu et al. (2023), and neural diarization models such as SOND for overlapped speech Du et al. (2022).

3.3. Large Language Models and Multi-Agent Paradigms

Recent work probes the multi-party capabilities of large language models. Systematic evaluations expose where ChatGPT and GPT-4 succeed and fail in multi-party reasoning Tan et al. (2023); Penzo et al. (2024), and zero- and few-shot LLMs are tested on dialogue act classification Martinenghi et al. (2024). Reasoning-oriented prompting often outperforms conventional fine-tuning for goal tracking and intent-slot recognition Addlesee et al. (2023), and hybrid intent recognition frameworks combine BERT-scale models with LLMs through uncertainty routing to balance accuracy and cost ópez et al. (2025). Reinforcement learning approaches optimize response coherence and negotiation outcomes, using utterance-level rewards Fan et al. (2024); Hiraoka et al. (2015) and cross-attention historical reward functions for self-supervised, context-aware generation Kiruluta et al. (2025); reward-decomposition strategies that mine latent logic further enable self-evolving reasoning Chen et al. (2026); while logical reasoning memory networks enhance QA accuracy Zhou et al. (2025). LLM-based explainable discourse parsing integrates contrastive learning to improve utterance-level accuracy Liu et al. (2025), and contrastive history-aware models further refine response consistency Hu et al. (2025).

The latest frontier extends MPD into multi-agent territory. MuPaS explores supervised fine-tuning for group dialogue generation with next-speaker prediction Wang et al. (2024); DICE-BENCH evaluates tool-calling in multi-round, multi-party scenarios with 1,607 synthesized dialogues Jang et al. (2025); SS-MPC encodes dialogue structure directly as sequential inputs within a Transformer encoder-decoder framework, removing the need for graph representations Jang et al. (2025); discourse-coherence response-guided context models (DRCR) improve generation on Ubuntu IRC Cao et al. (2026); context-aware turn-taking models predict when an agent should speak or stay silent Bhagtani et al. (2026); and MPCEval introduces a multi-dimensional benchmark for multi-party conversation quality Zhang et al. (2026). EverMemBench probes long-horizon collaborative memory across multi-group settings, revealing severe limitations in multi-hop attribution reasoning and temporal revision tracking by current LLM-based systems Hu et al. (2026). Synthetic data generation also emerges as a scalable resource for downstream MPD tasks: deterministic-constraint frameworks Penzo et al. (2026), tree-guided subspace partitioning for diverse data synthesis Wang et al. (2026), and knowledge-driven progressive-thought prompting for multi-turn dialogue augmentation Jiang et al. (2024) all help mitigate data scarcity. Orthogonally, parameter-efficient fine-tuning methods such as low-rank adaptation and its rotation- and shard-based variants make it practical to specialize large models for multi-party applications under limited compute Wang et al. (2024a) Wang, Chen, Jiang, Xue, Kong, and Wu; Wang et al. (2024b) Wang, Xue, Ye, Jiang, Chen, Kong, and Wu; Wang et al. (2025).

Takeaway. The trajectory in Figure 2 reveals a clear pattern: as model capacity grows, attention shifts from explicit structural modeling toward generative reasoning. Yet graph-based and symbolic approaches retain their value for capturing the relational structure that pure sequential models miss. Increasingly, state-of-the-art systems combine multiple paradigms, hinting at the neuro-symbolic hybrid architectures we identify as a key opportunity in Section 6.

4. Datasets

The diversity of MPD tasks is mirrored in the breadth of available corpora, which differ in source, scale, modality, and annotation depth. Table 2 summarizes representative datasets by category, providing a comparison of their scale and annotation focus. We organize them into four groups.

4.1. Structure and Disentanglement

Datasets in this category support research on threads, discourse relations, and conversational structure. The Asher et al. (2016, STAC Corpus), released at LREC 2016, contains multi-party chat dialogues annotated with full discourse structure based on Segmented Discourse Representation Theory (SDRT), including both linguistic and extra-linguistic interactions, enabling research on dialogue act classification and discourse-level reasoning. Li et al. (forthcoming, Molweni) Li, Liu, Kan, Zheng, Wang, Lei, Liu, and Qin, presented at COLING 2020, is an MRC dataset derived from the Ubuntu Chat Corpus consisting of 10,000 dialogues and 88,303 utterances annotated with discourse dependency structures based on a modified SDRT. Chang et al. (2023, MTDD) comprises 10,033 turns from 831 movies and TV series, annotated for conversational thread disentanglement and floor changes. Lerner et al. (2022, Bazinga!) consists of large-scale multi-party dialogues from TV and movie scripts with rich annotations for speaker diarization, addressee recognition, and entity linking, while Shaikh et al. (2010, OnlineChat-MPCorpus) focuses on online chatroom conversations annotated at four levels: conversational links, dialogue acts, local topics, and meso-level discourse structure.

4.2. Emotion and Social Relationship

TV transcripts dominate this category. Poria et al. (2019, MELD), released at ACL 2019, contains approximately 13,000 utterances from Friends, providing audio, visual, and textual modalities annotated with fine-grained emotion and sentiment labels. Its textual precursor Hsu et al. (2018, EmotionLines) includes 29,245 utterances from Friends and private Facebook chats annotated with seven emotion labels, and Zahiri and Choi (2017, EmoryNLP) utilizes Friends transcripts (S1–S4), comprising 12,606 utterances with seven crowdsourced labels. Chen et al. (2020, MPDD) supports Chinese-language affective MPD with annotations for emotions and interpersonal relationships, and Li et al. (2020, ALOHA (HLA-Chat)) links over a million dialogue lines from TV shows and movies to “Human-Level Attributes” (HLAs) based on audience-determined tropes for persona modeling.

4.3. Task-Oriented and Open-Domain

Gliwa et al. (2019, SAMSum) contains 16,369 messenger-like conversations with manually written third-person summaries, a primary benchmark for abstractive dialogue summarization, while Narayan et al. (2018, XSum) focuses on extreme summarization with 226,711 BBC article-summary pairs. Budzianowski et al. (2020, MultiWOZ) is a large-scale human-human dataset for task-oriented dialogue spanning seven domains, and the Lowe et al. (2015, Ubuntu Dialogue Corpus) contains nearly one million dialogues from technical support IRC logs for next-response selection. Manuvinakurike et al. (2021, IncTempSum) focuses on incremental abstractive summarization combining crowdsourced and expert annotations. More recently, Jang et al. (2025, DICE-BENCH) introduces 1,607 synthesized multi-round, multi-party dialogues for tool-use evaluation, and synthetic WMPC generation frameworks Penzo et al. (2026) demonstrate that instruction-tuned LLMs can produce structurally diverse MPD data under deterministic constraints such as dialogue structure and speaker stance, with turn-by-turn generation improving constraint adherence and linguistic variability. Sedoc et al. (2019, ChatEval) is designed for the evaluation of open-domain chatbots with manually filtered utterances and reference responses.

4.4. Multimodal and Naturalistic Meetings

The Carletta et al. (2006, AMI Meeting Corpus) offers 100 hours of meeting recordings (natural and elicited) with multimodal annotations including dialogue acts, head/hand gestures, and focus of attention. The Shriberg et al. (2004, ICSI Meeting Corpus (MRDA)) supplies 72 hours of naturally occurring professional meetings featuring over 180,000 hand-annotated dialogue-act tags. Yu et al. (2022, AliMeeting (M2MeT)) provides 120 hours of real-world Mandarin meetings recorded with 8-channel microphone arrays for diarization and multi-speaker ASR, while the Litman et al. (2016, Teams Corpus) captures realistic cooperative group dynamics across 63 teams through aligned audio, video, and detailed transcripts annotated for entrainment and dominance.

Despite this variety, current corpora remain skewed toward text-only English data, leaving multimodal, multilingual, and naturalistic professional settings comparatively underexplored, a gap that remains critical and we revisit in Section 6.

5. Evaluation

Evaluating MPD systems is harder than dyadic ones because quality spans multiple dimensions: semantic accuracy, structural coherence, social appropriateness, and collaborative effectiveness. Table 3 groups four metric families adopted in MPD studies, with representative metrics and typical use cases.

5.1. Automatic Generation Metrics

N-gram overlap measures including BLEU (BLEU-1 through BLEU-4) Bhatnagar et al. (2022); Fan et al. (2024); Li and Zhao (2023); Zhu et al. (2022), SacreBLEU Tan et al. (2023), and ROUGE (ROUGE-1, ROUGE-L) Bhatnagar et al. (2022); Chen and Metze (2012); Chen and Yang (2020); Li and Zhao (2023); Zhu et al. (2022) remain widely used for summarization and response generation. BLEU measures the precision of n-gram matches, while ROUGE emphasizes recall and is especially suitable for summarization. METEOR adds stemming, synonym matching, and word-order sensitivity Afantenos et al. (2015); Fan et al. (2024); Gu et al. (2022); Li and Zhao (2023). For semantic similarity, BERTScore computes overlap using contextual embeddings Bhatnagar et al. (2022); Ju et al. (2022), and other embedding-based metrics reduce reliance on lexical overlap Ju et al. (2022). Perplexity (PPL) gauges fluency by measuring the likelihood assigned to human-written references Ju et al. (2022); Laskowski (2010), though lower PPL does not guarantee relevance. Finally, Distinct-n (for example, Dist-1/2) quantifies diversity by computing the ratio of unique n-grams Ju et al. (2022), which is particularly relevant in MPD where repetitive answers harm conversational quality.

5.2. Classification and Selection Metrics

Accuracy, precision, recall, and F1 dominate addressee and intent classification, with variants including in-domain versus cross-domain accuracy Addlesee et al. (2023); Gatti de Bayser et al. (2019); Gu et al. (2021a) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang; Le et al. (2019); Mayfield et al. (2012); Sun et al. (2021), and macro-, micro-, and weighted-F1 providing nuanced evaluation under class imbalance Afantenos et al. (2015); Aina et al. (2019); Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang; Li et al. (2023); Liu et al. (2025); Meng et al. (2018); Wang et al. (2020). Ranking metrics such as Precision@k Gu et al. (2021a) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2023); Le et al. (2019), Recall@k Gu et al. (2021b) Gu, Tao, Ling, Xu, Geng, and Jiang; Gu et al. (2023), and Mean Reciprocal Rank Meng et al. (2018) evaluate candidate selection in retrieval-style tasks. Task-specific scores include ADR/RES for addressee-response pairs Ouchi and Tsuboi (2016), Exact Match (EM) for slot filling and structure prediction Li and Zhao (2021, 2023); Li et al. (2023); Zhou et al. (2025), AUC and PR-AUC for probabilistic classification Li et al. (2023); Park et al. (2022); Sia et al. (2022), and the recently introduced DICE-score for structured tool-calling Jang et al. (2025).

5.3. Structural and Discourse Metrics

Topic segmentation is assessed with $P_k$ and WindowDiff (WD) Galley et al. (2003); Purver et al. (2006), with lenient-match variants adapting them to noisy data Fernández et al. (2008). Discourse parsing employs graph-prediction F1 and relation-classification F1 Li et al. (2023) together with Link and Link&Rel metrics Liu et al. (2025) to assess discourse connectivity. Diarization Error Rate (DER) measures speaker attribution accuracy Du et al. (2022). Some works adopt multi-level evaluation, distinguishing sentence-, sequence-, and thread-level scores using Accuracy, Kappa agreement ($\kappa$), authoritativeness $r^2$, and Micro-averaged F-score Mayfield et al. (2012).

5.4. Human Evaluation

Despite advances in automatic scoring, human judgment remains indispensable for assessing fluency (grammatical correctness and naturalness) Fan et al. (2024); Gu et al. (2023); Ju et al. (2022); Liu et al. (2019), consistency and coherence (logical flow within context) Gu et al. (2023); Ju et al. (2022); Kiruluta et al. (2025); Liu et al. (2019), relevance (appropriateness to dialogue state) Fan et al. (2024); Gu et al. (2023); Liu et al. (2019), informativeness (usefulness of content) Fan et al. (2024); Gu et al. (2023); Liu et al. (2019), authority and trustworthiness (credibility) Mayfield et al. (2012), and interactional aspects such as step correctness, latency, and coherence Kiruluta et al. (2025). Task-specific measures further enrich the toolkit, including average reward in RL settings Hiraoka et al. (2015), cross-domain performance gap as a robustness measure Liu and Chen (2021), error reduction rate Laskowski et al. (2008), and composite scores that combine PPL, BLEU, Distinct, embedding similarity, accuracy, P/R/F1, and human ratings Ju et al. (2022); Wang et al. (2024).

Taken together, current evaluation practice remains fragmented: most studies rely on a handful of automatic metrics that imperfectly capture the social and collaborative dimensions of MPD. Section 6 argues for holistic, interaction-oriented benchmarks as a primary research opportunity.

6. Challenges and Opportunities

Building effective MPD systems requires confronting challenges distinct from dyadic dialogue. Figure 3 maps four bottlenecks to opportunities. We discuss each pair.

6.1. Combinatorial Complexity vs. Social Cognition

The transition from one-to-one to many-to-many interaction introduces a combinatorial explosion in conversational state. Models must track not only utterance history but also a dynamic graph of speakers, addressees, floor holders, and sub-threads. Current sequential LLMs flatten this structure, leading to failures in addressee prediction and contextually grounded turn-taking. The opportunity lies in architectures with inherent social cognition. Heterogeneous Graph Neural Networks naturally encode participants, utterances, and their evolving relationships as a dynamic graph. Multi-Agent Systems and Reinforcement Learning enable autonomous agents to learn emergent communication strategies through simulated interaction, moving beyond pattern imitation toward goal-driven social behavior.

6.2. The Structural-Semantic Divide vs. Neuro-Symbolic Fusion

In MPD, meaning is entangled with discourse position. Conversations comprise interleaved threads where a speaker may respond to an utterance several turns prior. While LLMs capture local coherence, they lack explicit mechanisms for global hierarchical structure, producing fluent but logically fragmented contributions. Neuro-symbolic hybrid architectures offer a promising path: an LLM augmented with a queryable discourse graph that tracks threads, stances, and decision states. This fusion would enable responses that are fluent and globally coherent, with explainable reasoning over conversational history.

6.3. The Evaluation Conundrum vs. Holistic Benchmarks

The field leans heavily on metrics borrowed from other NLP tasks, which are poor proxies for conversational quality in group settings. BLEU and F1 fail to capture pragmatic appropriateness, role fulfillment, consensus building, and group-level outcomes such as decision quality. The opportunity is to design holistic, multi-dimensional benchmarks that evaluate interdependent capabilities jointly (for example, response generation together with addressee prediction and topic tracking) alongside goal-oriented and human-in-the-loop protocols that measure functional contribution to group outcomes rather than surface mimicry. Progress in adjacent areas is instructive: trajectory-aware evaluation of research agents Chen et al. (2026) and large-scale benchmarks for dialogue understanding and generation Hu et al. (2026) show how multi-dimensional, process-level evaluation can move beyond single-score metrics.

6.4. The Data and Modality Bottleneck vs. Multimodal Grounding

Progress is constrained by the scarcity of large-scale, richly annotated MPD corpora, a problem compounded for non-English settings where dedicated benchmarks and datasets remain limited, as recent work on Cantonese capabilities and resources illustrates Jiang et al. (2025a) Jiang, Chen, Chen, Wang, Bao, Kong, Li, and Wu; Jiang et al. (2025b) Jiang, Truong, Chen, Bao, Wang, Chen, Wang, Kong, Li, and Wu. Most existing data is text-only, missing the gaze, gesture, prosody, and turn-yielding cues that humans use to manage group conversation. The path forward involves large-scale multimodal corpora and, more profoundly, deeper integration with computational social science methods. By learning to infer latent social variables such as leadership, facilitation, and group cohesion from conversational data, AI can evolve from a passive participant into an active, socially aware collaborator capable of adapting to pragmatic and cultural context.Recent advances in multi-modal proactive reasoning, which decouple perception from reasoning, point toward systems that can ground group interaction in visual as well as textual signals Zhou et al. (2025).

7. Conclusion

Surveyed Multi-Party Dialogue research across tasks, methods, datasets, and evaluation. The field advanced significantly, from statistical models through deep learning and graph reasoning to large language models and multi-agent paradigms. Yet gaps remain: approaches struggle with N-party interaction complexity, evaluation protocols inadequately capture social and collaborative quality, and data resources lag behind dyadic dialogue. The path forward calls for integrated modeling frameworks jointly addressing multiple MPD tasks, deeper grounding in external knowledge and social dynamics, principled integration of multi-agent reasoning with LLM capabilities, human-centered evaluation, and richer multimodal datasets covering diverse collaborative scenarios. Realizing these directions requires interdisciplinary collaboration spanning linguistics, sociology, psychology, and artificial intelligence. We hope this survey provides a clear synthesis of current progress and a practical roadmap for socially intelligent dialogue systems capable of supporting human-human and human-machine collaboration in complex environments.