I will focus on deep learning agent multi agent rag memory

## Innovative Design and Improvement Guidance for Agentic RL and LLM Reinforcement Learning

 LLMs are gradually evolving from single-turn Q&A machines into agentic systems capable of repeatedly interacting

 between reasoning and external tool use in multi-turn settings. From Search-R1 to ToolRL and SkyRL, models now need to

 not only think, but also search, calculate, call APIs, and continuously self-improve through RL across multi-

 step trajectories.

 ## 1. Innovative Design Improvements for Agentic RL Algorithms

 ### 1.1 Hierarchical Reinforcement Learning Architecture

 A hierarchical decision-making mechanism divides an Agents decisions into three levels: the strategic layer for task

 decomposition, the tactical layer for tool selection, and the execution layer for concrete operations. Each layer

 adopts a different RL policy.

 Automatic sub-goal discovery allows Agents to identify reusable intermediate sub-goals during training and construct a

 skill library.

 Automated curriculum learning emphasizes enabling Agents to progress autonomously from simple tasks to complex tasks

 without manually designed curricula.

 ### 1.2 Multimodal Environment Interaction