发现信息
- 作者:Uday Kamath, Vedant Vajre
- 年份:2026
- 发表源:CRC Press eBooks
- OpenAlex ID:https://openalex.org/W7214075423
- DOI:https://doi.org/10.1201/9781003668268-7
- arXiv:待核验
- 检索主题:memory-learning
- OpenAlex 引用数快照:0(2026-11-12)
摘要线索
Agentic AI moves reinforcement learning beyond the single-response setting of RLHF into environments where a model must observe, decide, act, and revise its beliefs over time. This chapter formalizes that shift by recasting LLM-based agents as policies operating in partially observable Markov decision processes, where the true state is hidden and success depends on memory, tool use, planning, perception, and long-horizon credit assignment. It develops the architectural foundations of agentic systems, including observe–think–act loops, synchronous and asynchronous execution, and the trade-offs between modular components and end-to-end policies. The chapter then examines how RL supports core agentic capabilities through systems such as RAP and VOYAGER for planning, ToolRL for external tool use, MemAgent for memory management, Reflexion for self-correction, and GRIT for active perception. A final customer-support case study integrates these ideas into a multi-turn agent that diagnoses complaints, invokes tools, remembers prior interactions, and improves behavior through reinforcement learning. Progressive capability integration is used to measure the marginal value of each component.
核验与筛选
- 当前状态:待筛选、去重和一手来源核验。
- OpenAlex 元数据和引用数只用于发现,不作为正式引文或质量判断。
- 记录正式版本、venue、方法细节、实验协议和局限后,才迁入稳定论文卡片。