Cohesive Conversations: Enhancing Authenticity in Multi-Agent Simulated Dialogues

Anonymous

TL;DR: We investigate the problems in multi-agent simulated dialogues over a span of time and propose a Screening, Diagnosis, Re-generation framework to instantly correct inconsistencies while bolstering multi-dialogue diversity.

Abstract: Recently, numerous studies have explored the idea of assemblies of autonomous agents driven by large language models as a society or collective group, where the agents interact with each other through text conversations. While individual dialogues appeared contextually appropriate when viewed in isolation, a wider examination of multiple interactions revealed a notable level of unnatural repetition and inconsistencies. This was particularly evident in recurring topics across dialogues, regardless of the distinct backgrounds and personas of the interacting agents. To address this problem, we propose a framework to automatically detect and rectify these unnatural dialogues and utterances. The proposed framework not only identifies inconsistencies and repetitive patterns but also corrects them to ensure a more coherent flow of conversations. Evaluations demonstrate the efficacy of our approach, as the post-correction dialogues exhibit a marked improvement in naturalness and variety. Through our study, we highlight the importance of viewing agent conversations holistically and present a solution that enhances the realism of multi-agent simulated interactions.

Paper Type: long

Research Area: Dialogue and Interactive Systems

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Reproduction study, Data analysis

Languages Studied: English