-
Background The paper discusses how LLMs like GPT-4 are key components in the progress towards Artificial General Intelligence (AGI). Despite their excellence in many tasks, integrating them into conversational agents that can maintain coherent, context-aware, and purpose-driven conversations is complex.
-
Existing Work Existing frameworks utilize the strengths of LLMs but have limitations in conversational settings and do not meet the comprehensive demands of more complex dialogues.
- Introduced the RAISE framework
-
Challenge 1: Establishing a dual-component memory system similar to the human brain to maintain coherence and context in conversation RAISE builds on the ReAct framework by adding scratchpad and retrieved examples, functioning as short-term and long-term memory features, to enhance the adaptability and controllability of agents.
-
Challenge 2: Building a complete agent construction scenario to ensure authenticity and relevance in real-world interactions RAISE improves agent performance in language processing, contextual awareness, and adaptability through a series of guided stages, including Conversation Selection, Scene Extraction, Chain of Thought Completion, and Scene Augmentation, leading to the LLMs Training phase.
-
The researchers conducted experimental evaluations on an in-house dataset focused on real estate sales, showing that RAISE surpasses traditional conversational agents in handling complex multi-turn conversations with better context adaptability and awareness. Although these experiments concentrate on the real estate domain, the RAISE framework's principles and methods are universally applicable and can be adapted to various domains.
The paper introduces the RAISE framework, which enhances the performance of LLMs in multi-turn dialogues, especially in real estate sales contexts, by incorporating an augmented memory system and a structured agent construction process.