Skip to content

๐Ÿ—ฃ๏ธ Dialogue Systems

๐Ÿ’ฌ EMNLP2026 ยท 21 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿ”ฌ ICLR2026 (10) ยท ๐Ÿ’ฌ ACL2026 (26) ยท ๐Ÿงช ICML2026 (5) ยท ๐Ÿค– AAAI2026 (5) ยท ๐Ÿง  NeurIPS2025 (8) ยท ๐Ÿงช ICML2025 (2)

๐Ÿ”ฅ Top topics: Dialogue ร—11 ยท LLM ร—6 ยท Sentiment Analysis ร—2

AI Psychosis: Does Conversational AI Amplify Delusion-Related Language?

This paper provides the first large-scale empirical study examining whether conversational AI amplifies delusion-related language during extended interactions, demonstrating that interactions with unconditioned conversational AI increase delusion-related language trajectories by an average of 233% in vulnerable simulated users, while conditioning AI responses on real-time inferred DelusionScore effectively reverses this harmful feedback loop.

Anticipating User Interpretations in Emotional Support Conversations via Inference-Time User-Side Simulation

Addressing the frequent conversational breakdown where sensible advice is misperceived by help-seekers as dismissive, judgmental, or pressuring, this paper presents Cognitive Look-ahead Simulation (CLS), an inference-time framework that infers structured user states, anticipates immediate user inner thoughts and state transitions via a lightweight cognitive world model, and selectively rewrites high-risk drafts before delivery, substantially reducing severe failure rates and improving multi-turn dialogue outcomes.

Beyond Single-Policy: Evaluating Composed Organization-Specific Policy Alignment in LLM Chatbots

Addressing the critical gap where real-world chatbot queries naturally span multiple organizational rules, COPAL introduces an automated evaluation framework driven by four empirical interaction patterns and target facets paired with explicit handling contracts, revealing an average composed-policy error rate of 33.1% across 9 frontier LLMs.

Client Simulation via Conversational Trajectory Modeling for Both Trainee Practice and Model Evaluation in Mental Health Counseling

PsyCLIENT tackles the overly compliant and behaviorally uniform nature of simulated counseling clients by decoupling static client profiles from dynamic conversational behavior trajectories, constructing the PsyCLIENT-CP benchmark of 120 Chinese clinical profiles and achieving superior authenticity and training utility.

From Chatbots to Confidants: A Cross-Cultural Study of LLM Adoption for Emotional Support

This paper presents the first large-scale cross-cultural empirical study of LLM adoption for emotional support across seven countries (\(N=4,641\)), disentangling individual demographics from country-level cultural variation via Cumulative Link Mixed Models (CLMM) and revealing a sharp divide between enthusiastic Anglosphere and skeptical Continental European users.

Know What You should not Know: Dynamic Attention Gating for Knowledge-Bounded Role-Playing

To tackle out-of-boundary knowledge leakage in role-playing conversational agents where Chain-of-Thought reasoning exposes anachronistic or cross-world knowledge, this paper proposes the Dynamic Attention Gating (DAG) Routerโ€”a lightweight network (<0.6% parameters of the host LLM) that keeps the LLM frozen and dynamically predicts per-layer FFN neuron gating masks from the system prompt, substantially boosting out-of-boundary compliance without compromising general conversational ability.

Learning an Efficient Multi-Turn Dialogue Evaluator from Multiple LLM Judges

Addressing the excessive inference overhead of multi-LLM judge ensembles and systematic biases in single judges, this paper proposes MTDEval, a lightweight evaluator trained on a new 10-dimensional multi-judge preference dataset (P2-MTD) using Thurstone's Case V model and Signal Detection Theory to jointly estimate judge reliabilities and dialogue quality, outperforming strong open-source baselines with over 10ร— faster inference.

Learning Diagnostic Clarifying Questions for Conversational Question Answering via Data-Efficient Dynamic Distillation

Addressing the dual challenges of non-convergent clarification and strict rollout filtering in open-domain conversational QA, D4CQ introduces procedural scaffolding, contrastive intent alignment, response-diagnosticity loss reweighting, and compatibility-gated hindsight relabeling, substantially outperforming existing clarify-or-answer baselines under a comparable interaction budget.

PersonaDual: Balancing Personalization and Objectivity via Adaptive Reasoning

To resolve the double-edged sword dilemma where personalized user personas enhance subjective alignment but degrade factual objectivity and logical reasoning, PersonaDual unifies general objective reasoning and personalized reasoning within a single LLM, adaptively switching between modes via a novel DualGRPO reinforcement learning algorithm.

Preference-Aligned Data Synthesis for Fine-Grained Intent Detection

Addressing inter-class ambiguity and intra-class homogeneity in LLM-synthesized data for fine-grained intent detection, this paper introduces a parameter-level generator alignment framework that mines multi-centroid prototype-aligned diverse positives and confusability-weighted cross-intent hard negatives for Direct Preference Optimization (DPO), delivering significant downstream gains in zero-shot and few-shot regimes while improving generator reliability.

REAT: A Reflective Experience-Augmented Tutoring Framework for Multi-turn Mathematical Instruction

To overcome the pedagogical rigidity and lack of systematic experience accumulation in LLM-based tutors during multi-turn interactions, REAT introduces an offline Observer-Critic-Mentor (OCM) distillation pipeline to extract problem-agnostic pedagogical experiences from historical dialogues, coupled with an online coarse-to-fine state-aware retrieval mechanism to dynamically deliver adaptive, scaffolded instruction.

Simulating Multimodal Multiparty Student Interactions with Large Language Models for Training Adaptive Pedagogical Agents

This paper introduces an LLM-based user simulation framework for triadic robot-student collaboration that serializes gaze and gestures into inline textual stage directions, using structural phase tagging and dynamic persona reinforcement to effectively train adaptive pedagogical reinforcement learning policies.

SpokenUS: A Spoken User Simulator for Task-Oriented Dialogue

To address the lack of realistic spoken user behaviors in text-based user simulators and the lack of proactive turn-taking in omni models, this paper introduces SpokenTODโ€”a 1,034-hour spoken TOD dataset augmented with four authentic speech behaviorsโ€”and SpokenUS, a grounded spoken user simulator featuring a dedicated turn-taking head for real-time proactive barge-in decisions.

Stranger, Fan, or Peer? A Systematic Study on the Role of Interlocutor in Persona-Based Dialogue Generation

This paper factorises interlocutor biography visibility into three independent stagesโ€”training, inference, and evaluationโ€”demonstrating that biographical copying is predominantly a training-time artefact rather than an inherent limitation of persona conditioning, while asymmetric disclosure creates an indirect leakage channel where target profile information surfaces through interlocutor turns.

UCO: A Multi-Turn Interactive Reinforcement Learning Method for Adaptive Teaching with Large Language Models

Addressing the critical flaw that existing LLM tutors optimize purely for answer correctness rather than genuine comprehension, this paper presents Unidirectional Cognitive Optimization (UCO), an interactive multi-turn reinforcement learning framework that leverages cognitive entropy reduction and dynamic ZPD scaffolding rewards to guide students through productive struggle while strictly mitigating solution leakage.

What Makes a Good Response? An Empirical Analysis of Quality in Qualitative Interviews

This paper introduces the Qualitative Interview Corpus spanning 14 real-world research projects and 16,940 participant responses, and uses automated LLM judges with linear mixed-effects modeling to show that relevance to key research questions and attributed meaning strongly predict contribution to research findings, whereas commonly used NLP proxy metrics like clarity and surprisal-based informativeness offer no predictive power.

When Humans and LLMs Disagree: Understanding Human-LLM Disagreement in Learner Intent Annotation for Educational Chatbot Dialogues

This paper introduces a Learning-to-Learn (L2L) theory-grounded hierarchical decision-tree annotation scheme for learner intent in educational chatbot interactions, revealing that while zero-shot LLMs exhibit agreement approaching human reliability (LLM \(\kappa = 0.67\) vs. human \(\kappa = 0.73\)), their alignment with human reference labels drops sharply (\(\kappa \le 0.52\)), and qualitative error auditing demonstrates that over half of disagreements stem from data ambiguity and human variability rather than model failure alone.

When Less Is More: An Empirical Study of Minimal Responses in Counseling Dialogues and the Behavior of LLMs

This paper presents a systematic cross-lingual empirical study showing that minimal responses (e.g., "Mm-hmm", "I see") are ubiquitous in real human counseling but virtually absent from synthetic datasets, and reveals that while state-of-the-art LLMs can generate minimal responses when explicitly instructed, they fail to deploy them spontaneously, while standard LLM-as-a-judge evaluation frameworks heavily penalize such essential non-intrusive listening behaviors.

When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues

This paper characterizes State Integrity Collapse (SIC) in long-horizon dialoguesโ€”a pervasive failure mode where LLMs accurately restate earlier constraints under probe yet fail to enforce them during generationโ€”demonstrating that 88.6% of multi-turn failures stem from generation-time state-binding breakdowns rather than memory loss, and introduces Structured State Anchoring (SSA) to recover up to 24.4 points in retention.

When the Bottleneck Shifts: Diagnosing and Closing Information Gaps in Interactive Decision-Making

This paper formalizes the Information Sufficiency Hypothesis (ISH) in interactive decision-making and Text-to-SQL, proposes a five-class taxonomy of information gaps, and develops PASCALโ€”a training-free acquisition-control data agent framework that prioritizes cost-normalized evidence gathering, active clarification, and adaptive retry breaker gating, allowing an on-premise 35B model to outperform frontier proprietary models by up to +14.7 pp on BIRD-Interact.

Who Am I? History-Aware Profiles for Student Simulation in Tutoring Dialogues

ProfileRL establishes history-conditioned student dialogue simulation by extracting structured five-dimensional cognitive-behavioral profiles, aligning the simulator via counterfactual DPO, and refining profile generation via GRPO with downstream simulation log-likelihood rewards.