GenAI Agents: Major Applications (Part1)
0 likes306 views
The document outlines the characteristics, types, and applications of generative AI (GenAI) agents, which are autonomous systems utilizing advanced generative AI models for interactive communication. It discusses the properties of individual and multi-agent systems, their roles in various domains such as software engineering, simulations, and their evaluation metrics. Key features include adaptability, task performance, and the ability to self-correct based on feedback.
GenAI Agents: Major Applications (Part1)
- 1. GenAI Agents: Major Applications (Part1) Vladimir Kanchev, PhD
- 2. Contents 1. Introduction. History and Basic Definitions. 2. Types and Structures of GenAI Agents. 3. Applications of GenAI Agents. 4. Evaluation of GenAI Agents. 5. Types of GenAI Frameworks and Technologies. 6. Challenges and Future of GenAI.
- 3. Introduction GenAI Agents Def: An autonomous system that uses advanced generative AI models to create and interact through a human-like text, images, audio, or other media. They often use text interfaces, enabling users to input prompts or queries and receive context-aware, conversational responses. (ChatGPT) a
- 4. Introduction General AI Agents Properties: ā autonomy: independently make decisions ā perception: gather information about their environment ā decision-making: select appropriate actions ā action: their actions change the state of their environment CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 5. History AI agents: ā¢ Basic concept of computations ā¢ Theory of mind ā¢ Society of Mind (Minsky) ā¢ Symbolic AI
- 6. History Types of AI agents: ā reflex agents: independently make a decision ā goal-based agents: gather information about their environment ā utility-based agents: select appropriate actions ā reinforcement learning (RL) agents ā GenAI agents ā AGI agents CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 7. History CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 8. Present WANG, Yuntao, et al. Large Model Agents: State-of-the-Art, Cooperation Paradigms, Security and Privacy, and Future Trends. arXiv preprint arXiv:2409.14457, 2024.
- 9. Basic Definitions The ChatGPT era ā November, 2022. Properties of LLM models: ā extensive knowledge base: the entire internet ā adaptability: context learning by zero-shot, few-shot learning ā human-computer communication ā scalability CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 10. Basic Definitions Disadvantages of LLM models: ā context length constraints ā prolonged knowledge update ā no direct tool support ā LLMās cannot employ code interpreters, calculators, etc. ā potential for biased or inaccurate output (hallucinations) ā dependence on the training of data CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 11. Contents 1. Introduction. History and Basic Definitions. 2. Types and Structures of GenAI Agents. 3. Applications of GenAI Agents. 4. Evaluation of GenAI Agents. 5. Types of GenAI Frameworks and Technologies. 6. Challenges and Future of GenAI.
- 12. Types of GenAI Agents Different types of GenAI depending on: ā number of LLM agents in the system: single or multi-agent ā the type of LLM model: single-modal or multi-modal ā the type of task specialization: general purpose and task- specific
- 13. Single GenAI Agents 5 major properties of a single GenAI agent: ā LLM model: serves as an agent brain, making decisions ā objective: terminal state the agent must achieve ā action: set of executable tasks that change the environment ā memory: stores the environment feedback and task history to improve performance ā reflection (rethink): evaluates previous actions and feedback CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 14. Single GenAI Agents 2 external parts of a single GenAI agent: ā tool: extensions of the agent's actions, such as calculators, code interpreters, or robotic arms. ā environment: provides input, constraints, and feedback that influence the agent's behavior and decision-making CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 15. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 16. Single GenAI Agents Planning Def: Building an action sequence based on set objectives and adapting it to the environment constraints to secure goal achievement. It is built on the reasoning of an LLM model not on a learned policy (RL agent). CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 17. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 18. Single GenAI Agents Memory in a GenAI agent: ā preserves knowledge and data from experience ā represents the current state of the GenAI agent ā can be textual, a vector/graph database, procedural ā supports adaptation ā facilitates personalization CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 19. Single GenAI Agents CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 20. Single GenAI Agents Rethinking the ability or retrospection in a GenAI agent: ā evaluates prior decisions ā enhances the GenAI agentās decision-making and learning ā improves the GenAI agentās adaptability CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 21. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 22. Single GenAI Agents Environment of the GenAI agent: ā is of a specific type ā is influenced by the agent itself ā helps dynamic adaptation ā helps interactive learning CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 23. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 24. Single GenAI Agents Action Def: An agent extends its functionality by using external tools, such as APIs, calculators, code interpreters, or specialized software, to perform complex actions beyond its internal reasoning capabilities. Thus, it interacts with its environment and achieves the objectives more efficiently. Types: single-tool, multi-tool, task-oriented, generalist, environment-specific tool users. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 25. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 26. Multi-agent Systems (MAS) Def: A collaborative framework of multiple interacting intelligent agents, each one with specialized roles or capabilities, working to achieve complex objectives. Their tasks usually span multiple domains, require distributed problem-solving, or have parallelized workflows. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 27. Multi-agent Systems (MAS) Types of multi-agent systems: ā multi-role coordination - cooperative, competitive, mixed, hierarchical ā planning Type - Centralized Planning Decentralized Execution (CPDE) and Decentralized Planning Decentralized Execution (DPDE) CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 28. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 29. Multi-agent Systems (MAS) CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 30. Multi-agent Systems (MAS) CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 31. CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 32. Contents 1. Introduction. History and Basic Definitions. 2. Types and Structures of GenAI Agents. 3. Applications of GenAI Agents. 4. Evaluation of GenAI Agents. 5. Types of GenAI Agent Frameworks and Technologies. 6. Challenges and Future of GenAI.
- 33. Applications of GenAI Agents Here we provide a few applications: ā a single task ā React/Reflexion agents, content/code creation, documentation/test cases generation ā multiple agents ā software development/maintenance, collaborative coding, world simulations CHENG, Yuheng, et al. Exploring Large Language Model based Intelligent Agents: Definitions, Methods, and Prospects. arXiv e-prints, 2024, arXiv: 2401.03428.
- 34. REACT Single GenAI Agent ā combines Reasoning and Acting ā iterative feedback loop: Reason Act Observe Refine ā ā ā ā core features: LLM-based reasoning, contextual memory, and tool integration (e.g. APIs, search engines) ā broad applicability: from customer support and healthcare to autonomous vehicles and R&D. YAO, Shunyu, et al. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv e-prints, 2022, arXiv: 2210.03629.
- 35. RIGAKI, Maria, et al. Out of the cage: How stochastic parrots win in cyber security environments. arXiv preprint arXiv:2308.12086, 2023.
- 36. YAO, Shunyu, et al. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv e-prints, 2022, arXiv: 2210.03629.
- 37. REFLEXION Single GenAI Agent ā learns through reflection ā self-corrects through feedback ā combines memory and adaptability. ā has real-world applications SHINN, Noah, et al. Reflexion: Language agents with verbal reinforcement learning.(2023). arXiv preprint cs.AI/2303.11366, 2023.
- 38. SHINN, Noah, et al. Reflexion: Language agents with verbal reinforcement learning.(2023). arXiv preprint cs.AI/2303.11366, 2023.
- 39. SHINN, Noah, et al. Reflexion: Language agents with verbal reinforcement learning.(2023). arXiv preprint cs.AI/2303.11366, 2023.
- 40. Agents for Software Engineering GenAI agents can be oriented to: ā a single SE task (single agent) ā code generation, code quality assurance (testing) ā end-to-end SE task (multiple agents) ā software development/maintenance
- 41. Agents for Software Engineering LIU, Junwei, et al. Large language model-based agents for software engineering: A survey. arXiv preprint arXiv:2409.02977, 2024.
- 42. Multi-agent Systems (MAS) End-to-end software development (SD) agent system: ā covers the whole SD life-cycle ā aligns with software process models: waterfall or agile ā has a role-specific specialization ā features a collaborative interaction ā relies on more advanced communication LIU, Junwei, et al. Large language model-based agents for software engineering: A survey. arXiv preprint arXiv:2409.02977, 2024.
- 43. Multi-agent SD System LIU, Junwei, et al. Large language model-based agents for software engineering: A survey. arXiv preprint arXiv:2409.02977, 2024.
- 44. Šulti-agent for World Simulation Here we have: ā scenario-specific simulations ā social and environmental interactions ā embodied agents with diverse roles ā advanced agent interactions ā controlled sandbox environments
- 45. MP5 ā A Multi-modal Open-ended Embodied System in Minecraft ā Minecraft ecosystem ā MineDojo framework ā multi-modal and open-ended tasks ā task handling: context dependent and process-dependent ā the MP5 System Architecture consists of: Parser, Percipient, Planner, Performer, Patroller QIN, Yiran, et al. Mp5: A multi-modal open-ended embodied system in minecraft via active perception. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. p. 16307-16316.
- 46. MP5 ā A multi-modal open-ended embedded system in Minecraft QIN, Yiran, et al. Mp5: A multi-modal open-ended embodied system in minecraft via active perception. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. p. 16307-16316.
- 47. Contents 1. Introduction. History and Basic Definitions. 2. Types and Structures of GenAI Agents. 3. Applications of GenAI Agents. 4. Evaluation of GenAI Agents. 5. Types of GenAI Frameworks and Technologies. 6. Challenges and Future of GenAI.
- 48. Evaluation of GenAI Agents A single GenAI agent evaluation is characterized by: ā¢ performance of narrow tasks ā¢ adaptability to perform unseen tasks ā¢ alignment with the userās goals ā¢ business metrics Metrics: Task-specific accuracy, reasoning and adaptability, user interaction quality, robustness and reliability, efficiency
- 49. Evaluation of GenAI Agents Task-oriented benchmarks for single GenAI agents: ā NLP-specific benchmarks (e.g. HELM) ā multi-modal benchmarks (e.g. SEED-Bench) ā agentic benchmarks (e.g. AgentBench) ā autonomous decision-making (e.g. ALFWorld)
- 50. Evaluation of GenAI Agents Human-AI interaction has: ā response quality (e.g. relevance, coherence, accuracy) ā user satisfaction and trust levels (user satisfaction score) ā efficiency in task completion (task completion time, reduction in user queries) ā ability to take feedback and improve over time (retrain latency)
- 51. Evaluation of GenAI Agents A multiple GenAI agents evaluation is oriented to: ā coordination and communication ā emergent behaviors ā scalability and resource efficiency ā business metrics
- 52. Evaluation of GenAI Agents Evaluation metrics of multiple GenAI agents for: ā coordination efficiency (inter-agent communication success rate) ā emergent behavior analysis (bias detection score) ā scalability and resource utilization (bandwidth efficiency) ā inter-agent communication (message latency) ā competitive and cooperative performance (cooperation score)