Tag
该论文提出 FlowSpec 规范驱动的方法，微调一个 90 亿参数模型，使语音客服代理能够遵循从未见过的受监管催收流程，合规检查率从 67–71% 提升至 93–100%。
该论文提出 PsyCIDRA，一个将自由形式精神科访谈与诊断推理相结合的双智能体框架，其中访谈智能体借助工具维护笔记、加载专家技能并检索 ICD-11 参考，诊断智能体则基于访谈记录生成假设并报告支持、冲突和缺失证据。模拟评估和针对 101 名人类参与者的盲测研究表明，该框架的诊断一致性显著优于直接提示。
ThuRunel is an advisory agent framework for two-phase high-stakes consultations (medical aesthetics, legal, education) that formalizes 'dynamic decoupling' — deciding what to ask, when to stop, what to resolve autonomously, and what to escalate to a specialist. Combining finite-state belief management, CoT teacher synthesis, and generation adapters, it outperforms eleven baselines and ships as a bilingual, source-citing web application.
This paper benchmarks the ability of Large Language Models (LLMs) to handle character attacks (ad hominem) in argumentative dialogues, revealing that LLMs prioritize logical defenses due to safety constraints, unlike human debaters.
The paper introduces Duplex-MPE, a benchmark for evaluating AI assistants in multi-party, full-duplex dialogues with 2,000 scenarios, assessing models on response initiation, accuracy, and silence preservation.
Introduces State-Grounded Conditioning (SGC), a design principle for user-facing LLM agents to condition on live state, reducing direction drift and improving latency and accuracy in in-game coaching scenarios.
The paper introduces a bilingual, multi-turn framework to study AI assistants' responses to repeated verbal abuse, distinguishing between hard disengagement and soft withdrawal, and presents experimental results with various API configurations showing heterogeneous behaviors.
This paper analyzes user-AI mistreatment in conversational systems by auditing LMSYS-Chat-1M conversations, revealing distinct patterns in hostility detection and its implications for AI safety and alignment.
TurnBench introduces a multi-domain benchmark for assessing turn-taking dynamics in spoken dialogue, featuring a hand-labeled corpus and standardized evaluation protocols for end-of-turn and interruption detection.
Introduces SWRouter, a similarity-contractive window router for multi-turn large language model conversations, addressing challenges in context segmentation and routing evaluation.
This paper proposes using semantic uncertainty derived from large language models to anticipate transition relevance places in spoken turn-taking, showing improved performance over baselines in dialogue systems.
This paper proposes a decoupled data approach to improve turn-taking in full-duplex dialogue by learning from real spoken dialogues while using text for semantics, leveraging a neural finite state machine framework to enhance naturalness and preserve semantic capabilities.
EmoStance is a method for empathetic response generation that uses emoji weak supervision to model response-side affective orientation, improving contextual specificity and perceived responsiveness in dialogues.
This paper proposes a multi-task learning approach that leverages turn-taking dynamics to enhance intent recognition in multi-party conversations, outperforming existing methods that ignore interaction patterns.
This paper proposes a generalized style-aware full-duplex framework with a lightweight turn controller LPS-TC, introduces a large-scale dataset WildTurn for real-world conversations, and presents a two-tier evaluation scheme to enhance proactive spoken interactions and response quality in dialogue systems.
ESCRAG-R1 is a unified framework that integrates retrieval-based psychological guidance into Group Relative Policy Optimization to enhance emotional support conversations, mitigating superficial splicing and achieving natural integration of professional guidance and empathy.
This paper introduces Real-TurnTurk, a multimodal Turkish conversational dataset for turn-taking prediction, featuring synchronized video, audio, and transcriptions. It uses genetic algorithms to optimize interpretable decision rules for predicting turn transitions in natural dialogue.
This paper introduces RO-PnR, a decision framework that learns when to ask clarifying questions versus providing corrections in multi-turn health misinformation interventions, achieving higher cost-adjusted utility with fewer turns.
This paper proposes a two-level hierarchical reinforcement learning framework called ToSCA for conversational agents, which uses DQN and PPO with a dual-granularity reward mechanism to improve strategy determination and response quality in multi-turn conversations.
This paper proposes a method to automatically generate diverse user personas using large language models for testing interview dialogue systems, reducing manual effort and increasing variation in simulated user behaviors.