dialogue state tracking

This work addresses the challenge that existing dialogue systems struggle to effectively establish and leverage common ground, often resulting in incoherent referential expressions in context-dependent conversations. To overcome this limitation, the authors propose a novel mechanism that integrates large language models with structured relational representations to explicitly model, store, and retrieve entities, events, and their interrelationships throughout a dialogue. This approach enables reasoning grounded in shared context and supports relational reference resolution. The designed memory architecture and dialogue policy not only facilitate clarification and confirmation but also dynamically maintain common ground over the course of interaction. Experimental results demonstrate that the proposed method significantly improves both the accuracy of common ground establishment and the consistency of subsequent referential utterances, outperforming current state-of-the-art baselines.

This work addresses the challenge of representational blur in existing dialogue systems, which often struggle to maintain a consistent shared context during long-term interactions due to reliance on purely textual representations. Inspired by human mental imagery, the authors propose an active visual scaffolding framework that incrementally externalizes dialogue states into persistent, traceable visual depictions. These concrete visual histories are integrated with propositional textual information to construct an explicit, multimodal common ground representation. The approach introduces, for the first time, a depictive intermediate representation analogous to mental imagery, effectively mitigating semantic flattening. Experiments on the IndiRef benchmark demonstrate that the proposed framework substantially outperforms full-dialogue reasoning, significantly reducing semantic ambiguity and enhancing contextual consistency, with the hybrid multimodal configuration achieving the best overall performance.

To address cognitive overload impeding knowledge integration and decision-making efficiency, this paper proposes a context-aware cognitive augmentation paradigm that transcends the passive response limitations of large language models (LLMs). Grounded in think-aloud studies conducted in exhibition settings, and integrating multimodal contextual modeling with cognitive state inference, we identify three core cognitive challenges: structural, retrieval-oriented, and application-oriented. Building upon these insights, we develop the first cognitive-state-driven LLM enhancement framework—unifying real-time contextual awareness, personalized reasoning assistance, and socially adaptive interaction, while enabling seamless transition from on-the-fly reasoning support to post-hoc knowledge organization. The framework establishes both theoretical foundations and practical design principles for human-centered, scalable AI systems that augment human cognition.

Current speech technologies (ASR/TTS) are largely confined to explicit semantic understanding, failing to model suprasemantic communicative signals—such as emotion, contextual dynamics, and implicit meaning—thereby limiting the naturalness and depth of human–machine interaction. To address this, we propose the “Beyond-Semantic Speech” (BoSS) paradigm and introduce a five-level spoken interaction capability framework (L1–L5) that systematically characterizes multidimensional implicit signals in speech. Integrating cognitive association theory with machine learning, we formalize the modeling of temporally dynamic, context-dependent non-semantic information. Empirical evaluation reveals that state-of-the-art spoken language models perform significantly below human baselines on BoSS tasks, exposing critical bottlenecks in situational awareness and interactional richness. This work establishes a novel paradigm, a principled framework, and a verifiable evaluation methodology for human-like speech intelligence.

In multi-agent AI communication, ambiguous speaker responsibility frequently causes role confusion, context drift, and unstable alignment, undermining interpretability and reliability. To address this, we propose a dual-mechanism framework: responsibility chain tracing and context integrity verification—enabling role-aware modeling, continuous responsibility flow propagation, and consistent context maintenance. We introduce structured evaluation metrics—pragmatic consistency, responsibility flow completeness rate, and context stability—and develop a lightweight G-Code configuration language with modular APIs to support high-context human-AI dialogue modeling. Our approach integrates role modeling, state tracking, and structured validation, leveraging human annotation, automated scoring, and bootstrapped statistical analysis. Experiments demonstrate a 42% improvement in responsibility flow completeness and a 37% gain in context stability, significantly enhancing structural robustness in multi-turn interactions—without reliance on affective signals or surface-level heuristics.

This study addresses the challenge of capturing and sharing users’ cognitive traces in complex knowledge work, which often leaves AI systems without a meaningful understanding of authentic task context. The authors propose a novel paradigm centered on in-situ, user-generated snippet memoing—fragmentary notes that serve as primary sources of contextual information for AI. For the first time, this approach treats such cognitive traces as first-class inputs and grants users direct editorial control over their contextual representations. Through a probe system comprising interaction mechanisms and a context-organizing interface, users can instantly record, inspect, and restructure their thought processes, thereby guiding AI to produce responses more aligned with their intent. Experimental results demonstrate that this method significantly enhances users’ task awareness and ability to structure thinking, with 78.1% of participants preferring AI responses grounded in snippet-based context, leading to greater perceived agency and ownership in collaborative creation.

This work addresses the challenge of maintaining coherence in large language models during nonlinear, hierarchical, and multi-branch human conversations, where inefficient context utilization often leads to degraded performance. To this end, the authors propose Context-Agent, a dynamic discourse tree framework that models multi-turn dialogues as an expandable tree structure, aligning with the inherent nonlinearity of natural conversation and enabling effective context maintenance and navigation across topical branches. Additionally, they introduce NTM, the first benchmark specifically designed for evaluating nonlinear long-range dialogue systems, thereby moving beyond conventional linear modeling paradigms. Experimental results demonstrate that the proposed approach significantly improves task completion rates and token efficiency across multiple large language models, substantiating the effectiveness of structured context management in complex, dynamic conversational settings.

Existing approaches to multimodal dialogue understanding often fail to explicitly model the dependency between contextual history and the current utterance. This work proposes CUCI-Net, which, for the first time, formulates context-utterance dependencies as interpretable cues. By employing structured encoding to differentiate contextual and utterance representations, the model integrates local modality-specific evidence with global contextual evidence to generate guiding cues. These cues then drive a cue-guided multimodal interaction mechanism. Evaluated on mainstream multimodal dialogue benchmark datasets, CUCI-Net significantly outperforms existing state-of-the-art models, demonstrating both its effectiveness and methodological novelty.

This work addresses the challenge of modeling and naturally expressing implicit chains of thought in full-duplex dialogue by proposing a Graph-of-Thought (GoT) architecture. The approach employs a multi-level perceptual framework to capture the causal and temporal dependencies from communicative intent to verbal behavior, integrating a hierarchical annotation scheme, graph-structured reasoning, and a Transformer backbone to enable streaming dynamic inference on high-quality, human-annotated, controllable dialogue corpora. The study introduces the first foundational model tailored for full-duplex dialogue behavior modeling, demonstrating robust behavior detection and interpretable chain-of-thought generation on both synthetic and real-world data, thereby establishing a new benchmark for dialogue reasoning.

Traditional asynchronous, data-driven explanations often fail to effectively convey information due to their lack of audience adaptation and interactivity. To address this limitation, this work proposes the CODEX framework, which introduces autonomous agents into the data explanation process for the first time. By integrating multimodal generation, dynamic audience modeling, and context-aware decision-making mechanisms, CODEX constructs adaptive information interfaces in real time. The framework continuously evaluates interaction progress against predefined communicative intents and dynamically adjusts its explanatory strategies to deliver personalized and highly effective data interpretations. This study establishes both theoretical foundations and methodological support for interactive, context-sensitive, data-driven explanation systems.