發展英語作為外語（EFL）的口說能力，通常面臨缺乏練習對象以及缺乏真實且具意義情境的困境。雖然人工智慧（AI）聊天機器人已被視為潛在的解決方案，但先前研究指出，現有的對話式 AI 系統在彈性、持續性、適應性及具意義的回饋方面仍有侷限，而這些要素對於有效溝通至關重要。通用的大型語言模型（LLM）通常難以維持結構化、目標導向的對話，或提供具針對性的教學鷹架。為了能解決上述限制，本研究提出多模態對話式 AI 系統，該系統將「多模態對話回應理論」（CRT）與四大智慧機制相結合，包含：彈性（容錯性）、持續性（透過圖像辨識進行多模態情境化）、適應性（動態難度調整）及回饋（隨選診斷性評量）。本系統架構扮演「能力更強的他者」（MKO）的角色，旨在提供高度情境化與個人化的 EFL 學習體驗，使學生能保持在「近側發展區間」（ZPD）內。為了檢驗所提方法之成效，本研究進行為期五週的準實驗研究，研究對象為 46 名大學四年級學生，並分為兩組：實驗組（EG）使用本研究提出之多模態系統，對照組（CG）則使用通用的純文字 AI。研究結果顯示，實驗組在整體英語能力的表現上優於對照組，特別是在「連貫性與流暢度」以及「文法多樣性與準確性」方面。此外，到了第三週，實驗組在任務表現上的「歐洲語言共同參考架構」（CEFR）等級提升幅度也高於對照組。至關重要的是，研究結果揭示「AI 輔助下的任務表現」與「獨立學習成效」之間存在脫鉤現象。雖然多模態的互動參與成功維持即時對話，但對系統生成之關鍵字的過度依賴，導致低英語能力的學習者產生「鷹架依賴」。歸根結底，長期的學習成效取決於主動的語言產出與自我糾正，而非僅僅取決於互動頻率。本研究還發現「自我調節陷阱」：被動閱讀錯誤筆記反而對期末測驗分數產生負面影響。儘管該系統作為「能力更強的他者」能有效降低口說焦慮，使用者仍感受到不同程度的技術摩擦與評量壓力。這些發現表明，只要學習者能主動運用回饋來修正錯誤，而非僅是被動閱讀，將對話回應理論與智慧機制整合至 AI 代理人中，便能有效促進永續的英語外語（EFL）對話學習。;Developing speaking skills in English as a Foreign Language (EFL) often suffers from a lack of speaking partners and authentic， meaningful contexts. While artificial intelligence (AI) chatbots have emerged as potential solutions， previous studies indicate that current conversational AI systems remain limited in flexibility， sustainability， adaptability， and meaningful feedback， which are essential for effective communication. General Large Language Models (LLMs) often struggle to maintain structured， goal-driven dialogue or provide targeted pedagogical scaffolding. To address these limitations， this study proposes a multimodal conversational AI system that integrates Multimodal Conversation Response Theory (CRT) with four smart mechanisms: flexibility (error tolerance)， sustainability (multimodal contextualization via image recognition)， adaptability (dynamic difficulty adjustment)， and feedback (on-demand diagnostic assessments). Acting as a More Knowledgeable Other (MKO)， this framework aims to provide a highly contextualized and personalized EFL learning experience that keeps students in their Zone of Proximal Development (ZPD). To examine the effectiveness of the proposed approaches， a quasi-experimental study was conducted over five weeks with 46 university seniors， divided into an Experimental Group (EG) using the proposed multimodal system and a Control Group (CG) using a baseline text-only AI. The results revealed that the experimental group outperformed the control group in overall proficiency， particularly in coherence and fluency， and grammatical range and accuracy. Furthermore， the experimental group achieved higher Common European Framework of Reference level gains in task performance by the third week. Crucially， findings revealed a disconnect between AI-supported task performance and independent learning achievement. While multimodal engagement successfully sustained real time dialogue， an over reliance on system generated keywords created a scaffolding dependence for lower proficiency learners. Ultimately， long-term learning achievement is driven by active text production and self-correction rather than mere interaction frequency. The study also identified a passive feedback-browsing behavior， where passively reading mistake notes negatively impacted final exam scores. Although the system functioned effectively as a More Knowledgeable Other to lower speaking anxiety， users still experienced varying degrees of technical friction and evaluative pressure. These findings suggest that integrating Conversation Response Theory and smart mechanisms into artificial intelligence agents can effectively facilitate sustainable English as a Foreign Language conversation， as long as learners actively apply the feedback to correct their mistakes instead of just passively reading it.
