arXiv cs.CL
·
6 hours ago
A study of 24 participants interacting with a memory-augmented conversational AI over 10 sessions found that conversational quality affects immediate enjoyment but not future sessions, while perceived memory influences later enjoyment through increased self-disclosure. Relational turning points—discrete moments of improvement or decline—were identifiable in multimodal behavioral data, with enjoyment improvements persisting more reliably than enjoyment declines recovering. The research suggests human-AI relationships develop through both gradual accumulation and sudden shifts in interaction quality.
arXiv cs.CL
·
6 hours ago
Researchers analyzed over 4,500 writing sessions with Penny, an LLM-powered chatbot designed to provide feedback on English writing by Japanese EFL learners. The study found that learners follow two main interaction patterns: a Revision Loop where feedback leads to error correction, and a Chat Loop involving sustained dialogue, with high-proficiency learners engaging in more open negotiation while low-proficiency learners rely more on repetitive corrections. The results suggest chatbot design should be differentiated by proficiency level to promote deeper cognitive engagement rather than focusing solely on error correction.
arXiv cs.CL
·
6 hours ago
Researchers introduced DialogueVPR, a conversational approach to visual place recognition that uses interactive dialogue instead of static image-to-location matching. The method includes DQ-pilot, an intelligent questioner trained via supervised fine-tuning on 20,000 curated dialogue examples followed by reinforcement learning on 10,000 harder examples. This dialogue-based reasoning approach significantly outperforms traditional single-shot retrieval methods for handling ambiguous and incomplete natural language descriptions of locations.
arXiv cs.CL
·
6 hours ago
Researchers developed a dialogue summarization framework that models both semantic content and emotion dynamics using a hierarchical Chain-of-Agents approach that decomposes conversations by topic segments and participant-specific utterances. The method was evaluated on multimodal dialogue datasets using small language models and introduces new emotion trajectory metrics to measure how well summaries preserve emotional flow across conversations. The framework enables more accurate dialogue summarization that captures both informational content and emotional progression, moving beyond existing approaches that treat dialogue like monologic text.
arXiv cs.CL
·
6 hours ago
Researchers introduced MAPS, a framework that enables AI dialogue systems to model conversations between agents with distinct cognitive perspectives while still reaching shared understanding. The framework uses domain-weighted profiles, GRU-based memory, and token-level attention mechanisms, and was evaluated on three dialogue datasets including EmpatheticDialogues and MultiWOZ. The approach allows dialogue systems to maintain individual reasoning styles while achieving semantic alignment, rather than forcing uniformity across all participants.
