Multi-Agent Reflection Loop: A Multimodal Framework for Cross-Cultural Narrative Adaptation

ACL ARR 2026 January Submission9452 Authors

Keywords: Cultural Adaptation, Cross-Cultural Learning, Multi-Agent Systems, Multimodal Interaction, Narrative Consistency

Abstract: Cross-cultural narrative understanding demands language models to not only excel at text generation, but also perceive users’ implicit cultural cognitive states and dynamically align with them. However, existing interactive narrative systems predominantly rely on static preset scripts or monolithic large language models (LLMs). These approaches fail to resolve semantic conflicts arising from multimodal user inputs (text, behavior, emotion), leading to feedback lacking cultural adaptability and even the generation of contextual illusions. To address these limitations, this paper proposes a Multimodal Culture-Aware Multi-Agent System (MC-MAS), which achieves high-precision cross-cultural narrative adaptation through a collaborative agent mechanism. Specifically, we design three specialized functional agents (behavioral, linguistic, cultural) to process heterogeneous user signals, and introduce a core coordinator agent. The coordinator employs a novel “reflection-reconstruction” loop mechanism, which can automatically detect cross-modal consistency conflicts and iteratively optimize narrative generation strategies. We validate the MC-MAS framework in a narrative scenario rooted in the exile literature of the Weimar Republic. Experimental results demonstrate that, compared with static models and single LLMs, our method significantly enhances the accuracy of cultural context alignment while preserving narrative coherence, and effectively alleviates users’ cognitive load.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: chain-of-thought,LLM/AI agents

Contribution Types: NLP engineering experiment

Languages Studied: China

Submission Number: 9452