Source-linked AI summary
CGMI: Configurable General Multi-Agent Interaction Framework
Shi Jinxin, Zhao Jiabao, Wang Yilei, Wu Xingjiao, Li Jiawen, He Liang
TL;DR
LLM-based agents can produce superficial content because of limited domain expertise and inadequate cognitive architecture, while also struggling with stable personas and realistic coordination. CGMI addresses these issues through tree-structured persona management, an ACT*-based skill-library architecture, and general agents; classroom experiments reported interactions aligned with actual teaching and supported instructional analysis.
Problem
LLM-based agents generate superficial content and struggle with domain expertise, effective cognition, stable role behavior, and realistic multi-agent coordination.
Method
CGMI combines tree-structured persona assignment, detection, and maintenance with an ACT*-inspired cognitive architecture using memory, reflection, planning, and domain skill retrieval.
Results
Classroom simulations with teacher, student, assistant, and supervisory agents produced interactions reported as aligned with actual teaching.
Takeaways & Limitations
CGMI supports domain-specific social simulation and can assist teacher instruction, evaluate teaching competencies, and validate teaching hypotheses.
Abstract
from arXiv · showhide
Benefiting from the powerful capabilities of large language models (LLMs), agents based on LLMs have shown the potential to address domain-specific tasks and emulate human behaviors. However, the content generated by these agents remains somewhat superficial, owing to their limited domain expertise and the absence of an effective cognitive architecture. To address this, we present the Configurable General Multi-Agent Interaction (CGMI) framework, designed to replicate human interactions in real-world scenarios. Specifically, we propose a tree-structured methodology for the assignment, detection, and maintenance of agent personality. Additionally, we designed a cognitive architecture equipped with a skill library based on the ACT* model, which contains memory, reflection, and planning modules. We have also integrated general agents to augment the virtual environment's realism. Using the CGMI framework, we simulated numerous classroom interactions between teacher and students. The experiments indicate that aspects such as the teaching methodology, curriculum, and student performance closely mirror real classroom settings. We will open source our work.
Introduction
CGMI addresses superficial LLM-agent behavior, unstable personas, and coordination difficulties by combining structured personas, cognitive mechanisms, and configurable multi-agent interaction. Its classroom simulation produced interactions reported as aligned with actual teaching.
- LLM-based agents can generate superficial content, forget character settings, and behave inconsistently during multi-turn interactions.The paper identifies limited domain expertise, missing cognitive architecture, context-window constraints, and role instability as central challenges.
- The cognitive architecture uses declarative and procedural memories with reflection and planning to improve agents’ handling of complex tasks.It is inspired by ACT* and uses CoT and CoA processes to extract memories and retrieve skills.
- The tree-structured persona model assigns, detects, and maintains agent traits, reducing memory pressure and improving behavioral stability.It addresses the difficulty of preserving comprehensive role settings in long, multi-turn dialogues.
- In a classroom scenario, CGMI simulated teacher, student, assistant, and supervisory agents, and the resulting interactions aligned with actual teaching.The reported uses include assisting teacher instruction, evaluating teaching competencies, and validating teaching hypotheses.
- CGMI combines a skill-library cognitive architecture, tree-structured persona management, and configurable general agents for realistic multi-agent interaction.The framework targets complex problem solving, stable role behavior, and scenario-specific collaboration.
Related Work
Prior work applies multi-agent systems to domain-specific tasks and to simulations of human social behavior. The supplied related work also includes a Big Five personality representation organized as a hierarchical scale.
- Domain-specific agent systems: Research has used multi-agent systems for healthcare recommendations, software development coordination, and customized educational support.Examples include communication feedback for treatment recommendations, coordinated design and coding workflows, and instructional decision support.
- Human-interaction simulation: Other studies construct agents and environments that emulate human activities, social roles, situational learning, and educational interactions.The examples span simulated towns, social-role communication, online visual-knowledge acquisition, and GPT-based education research.
- Personality representation: The Big Five personality scale is represented as a hierarchy with five coarse personalities and five fine-grained leaves under each coarse personality.Coarse dimensions range from 5-25, leaf dimensions from 1-5, and larger values indicate more pronounced characteristics.
Method
CGMI combines structured persona management, an ACT*-inspired cognitive architecture with a configurable skill library, and scenario-specific general agents to simulate classroom interactions. Its classroom implementation coordinates planning, staged teaching, supervision, personality consistency, and student participation.
- Tree-Structured Persona Model: The persona model uses Big Five, teaching-style, and learning-style scales organized as a scored tree traversed with depth-first search.Each node contains a description and score; the tree supports personality assignment, detection, and maintenance.
- Tree-Structured Persona Model: Randomized multi-level testing checks coarse- and fine-grained traits, restoring selected values when errors reveal incomplete personality memory.The procedure is intended to improve efficiency and conserve contextual-window resources while detecting fine-grained inconsistencies.
- Configurable General Multi-Agent Interaction Framework: The classroom scenario uses teacher and student role agents alongside assistants and supervisors that plan lessons, monitor stage transitions, maintain persona consistency, and select students based on willingness to speak.The teaching assistant distributes plans, the supervisor uses the plan and classroom state to control progression, and the consistency checker aligns statements and participation with personality and classroom status.
- Cognitive architecture equipped with skill library: The cognitive architecture comprises working, declarative, and procedural memory plus a configurable skill library that supplies domain knowledge for reflection and planning.Working memory receives external information and refines it through Chain of Thought and Chain of Action processes into declarative and procedural knowledge.
- Framework construction: CGMI combines a tree-structured persona model, skill-library cognitive architecture, and configurable multi-agent interaction framework for scenario-specific social simulation.The framework assigns, detects, and maintains agent traits while supporting memory, reflection, planning, and coordinated interaction.
Experiments
Experiments evaluated CGMI in virtual classroom teaching through FIAS analysis and comparative studies of cognitive architecture, personality modeling, and interaction logic. The results indicate richer, teacher-guided interactions, more consistent personalized behavior, and less random student selection.
- Analysis of Teaching Behavior: FIAS analysis found rich virtual-classroom interactions consistent with actual teaching across three mathematics sessions.Two trained encoders independently coded the sessions, and their evaluations were consistent.
- Analysis of Teaching Behavior: 61.23% of discourse came from teacher behaviors, while teacher-prompted student behaviors represented 23.53% and student-initiated interactions about 15.23%.The indirect-to-direct teacher influence ratio remained below 1, indicating teacher-led classroom control alongside student engagement.
- Cognitive Architecture: The cognitive architecture used student feedback and teaching-skill retrieval to support reflection and planning within and across lessons.Mrs. Smith adapted encouragement, gamified instruction, and later reassurance based on reflections about Emily’s anxiety and progress.
- Tree-Structured Persona Model: Personality allocation made student expressions more nuanced, while Emily maintained a consistent persona across classroom interactions.Without assigned traits, five students showed notably uniform expressions; assigned traits produced more individualized responses.
- Analysis of Interaction Logic: The willingness judgment agent recommended John 4, Emily 9, Ryan 6, Samantha 1, and Ying Zheng 8 times, unlike random selection.Its recommendations considered personality, classroom dynamics, and subject knowledge, whereas random selection recommended John 7, Emily 3, Ryan 4, Samantha 6, and Ying Zheng 8 times.
- Analysis of Interaction Logic: Teaching plans and student-state analysis enabled targeted questioning, such as asking Ying Zheng to explore different solutions.CGMI therefore supported interaction choices based on role settings and classroom circumstances rather than random selection.
Conclusion
The paper concludes that CGMI combines personalized multi-agent interaction, a skill-library cognitive architecture, and general agents to simulate domain-specific social phenomena. Its classroom case study verified the framework through analyses of teaching interactions and interaction logic, while future work targets experimental data and teaching applications.
- Conclusion: CGMI supports personalized, anthropomorphic interactions and collaborations among multiple agents for simulating domain-specific social phenomena.The framework combines configurable agents with general agents to enhance scenario authenticity.
- Conclusion: Its cognitive architecture combines domain knowledge with reflection and planning while condensing working memory into declarative and procedural memories.The framework also uses general agents to further enhance the authenticity of simulated scenarios.
- Conclusion: The classroom teaching scenario verified CGMI through comparative analyses of interaction content and logic between teachers and students.The conclusion presents the classroom simulation as evidence for the framework’s effectiveness.
- Future Work: Future applications include generating social experimental data for large-model development and supporting teaching and gamified teaching.These uses are stated as future directions rather than demonstrated outcomes.
Appendix
The appendix documents the character settings and tree-structured scales used to configure the simulated agents.
- Appendix: The appendix presents character settings for each character, a tree-structured learning-style scale, and a teaching-style scale.
Role Set
Role initialization configures agents using career, identity, background, personality, and teaching or learning style information.
- Role Set: Role agents are initialized with career, name, basic information, personality, and teaching or learning styles.The appendix identifies separate character-setting figures for Mrs Smith and five students.
Sternberg Thinking Styles in Teaching
Mrs. Smith’s teaching style is represented with the Sternberg Thinking Styles in Teaching Inventory as a scored tree structure, where higher scores indicate closer matches to her observed style.
- The inventory represents Mrs. Smith’s teaching style as a tree, with Level-2 scores measuring description-to-style matches from 1 to 7.Each Level-1 score sums the scores of its child nodes.
Solomon’s Learning Styles
Students’ learning styles are represented with Solomon’s Learning Styles Inventory as a tree-structured, four-dimensional classification based on selections across sub-nodes.
- Solomon’s inventory classifies students across four learning-style dimensions using Level-1 category types.Level-1 nodes represent the resulting type in each dimension.
- Each Level-2 node records the description and choice associated with an evaluation question.
- The category is selected by comparing how often students choose alternatives a and b across 11 sub-nodes.The more frequently selected alternative determines the represented category.