Introduction to LLM agents

As large language models (LLMs) gain increasing attention for real-world applications, research is increasingly focusing on automating workflows through the use of agents. This presentation provides an overview of fundamental concepts related to autonomous agents and introduces a line of research that leverages LLMs for designing agents capable of real-world tasks. We discuss both classical and modern approaches to agent design and compare them to illustrate how LLMs can enhance agents' decision-making processes. Finally, we present our research on an agent system where agents engage in conversations, with their self-emotions being simulated through a series of events. Additionally, we report how these simulated emotions influence the agents' dialogue behaviors in both dyadic conversations and group discussions.

Transcript

1. Introduction to LLM agents Couger AI
2. Autonomous Agent • Computational systems that sense and act autonomously the complex dynamics in the enviroment, and by doing so realize a set of goals or tasks for which they are designed.
3. What are agents for? • Autonomous agents are recognized as a symbol of AGI. • accomplish tasks through self- directed planning and actions.
4. Challenges and the LLM solution • Enviroment representations. • World knowledge comprehension. • Complexity of human-like decision making. • Appropriate actions/human-like behaviors selection.
5. LLM-driven Agent Architecture
6. Agent Profile • The information for the role of the agent. • age • gender • career • Generation method • Handicrafting • LLM generation • Real-world human simulation
7. Memory • Type • short-term: dynamics that is important for understanding current state and making decisions. • current conversation history • scene description • long-term: information that are important over time. • past decision-making trajectories • presets of behaviors
8. Decision-making • Reasoning • Single-path • Chain of thought • Multi-path • Tree of thoughs • Feedback Planning • Environment feedback • objective in the virtual world • Human feedback • human supervision
9. Action • Action space • Leveraging knowledge • plan • converse • Using external tools • call APIs • search knowledge base
10. Example - Recommendation RecMind
11. Recommendation • Consider both customer information and common information • Decompose tasks and select the best reasoning path • Execute tasks with approprite actions
12. Example - World Simulation Generative Agents
13. World Simulation • Streaming memory
14. • Make the plan • Conversation • Idea spreading 14/33 Emergent Communications
15. Example - Role playing
16. Example - Self emotion simulation • Responses are influenced by self-emotions. 16/33
17. Agents and self-emotion simulation[6] • Creation of schedules based on profiles. • Generation of random events. • Analysis of self-emotions. • Communication with self- emotions. 17/33 [6]Zhang, Q., Naradowsky, J., & Miyao, Y. (2024). Self-Emotion Blended Dialogue Generation in Social Simulation Agents. arXiv preprint arXiv:2408.01633.
18. Analysis • Positive self-emotions lead to more optimistic startegies and negative self-emotions lead to more pessimistic ones. 18/33
19. Self emotion - group discussion • The self-emotion of a member could influence the disussion results of a group. 19/33
20. Analysis • Agent with self-emotion incorporated behaves differently. 20/33
21. Analysis • Agents with negative self-emotions express more personal opinions, leading to more disagreements. • Positive self-emotions lead to quick agreements and focus on details. 21/33
22. A typical workflow in real world • Human supervision can be involved in the interaction loop to improve the performance.
23. Conclusion • LLM agents are good approach to • real-world problem representation and simulation • tasks with complicated procedures and external knowledge • applications that involves self-learning and self-improvement