Dialogizer: Context-aware Conversational-QA Dataset Generation from Textual Sources
https:// CGMI: Configurable General Multi-Agent Interaction Framework
https://arxiv.org/abs/2308.12503
“With the capabilities of large language models (LLMs) such as GPT4 (OpenAI 2023), we can construct more complex environment and create more realistic agents to simulate social phenomena. However, when using LLMs to complete ABSS tasks, the following issues need to be addressed: (1) How to trigger the capabilities of LLMs to solve complex problems? (2) How to ensure that agents have a stable role and behavior output based on LLMs without forgetting? (3) How to design a communication mechanism for LLMs-based agents to truly simulate interactions? Existing LLMs-based agents are mainly divided into action agents (Yao et al. 2023; Press et al. 2023) and plan-and execute agents (Wang et al. 2023a). Action agents make decisions based on previous outputs and are suitable for small tasks. Plan-and-execute agents formulate and execute action plans, suitable for long-term goal tasks. However, in complex scenarios, LLMs-based agents may produce mechanical and superficial content or not execute according to the plan. Inspired by the Adaptive Control of Thought (ACT*) model (Anderson and R 1983), we designed a cognitive architecture equipped with skill library for agents. Specifically, we employ the Chain of Thought (CoT) and Chain of Action (CoA) methods to extract declarative and procedural memories from the agent’s working memory. During the reflection and planning processes, content is retrieved from the skill library, ensuring deeper and more specialized insights.
Assigning each intelligent agent with a unique identity, personality, and capability (Wang et al. 2023c) can offer a more humanized and emotional interactive experience, and also enhance the realism of simulating complex social scenarios (Argyle et al. 2023). Although LLMs like GPT4 possess strong role-playing capabilities, we found that LLMs tend to forget the original character settings in multi-turn dialogues and make decisions that are inconsistent with the character’s design. Additionally, due to the limitations of the context window, it’s challenging to set roles comprehensively and in fine detail. To address these issues, this paper introduces a tree-structured persona model for character assignment, detection, and maintenance, which is beneficial for agent interaction performance.”