Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Chapter 10 Agentic-RL: Agent Reinforcement Learning Training

📖 "If Prompt Engineering is writing a 'user manual' for an Agent, then Agentic-RL is letting the Agent figure out the optimal approach through repeated practice."

🎓 Learning Objectives

After completing this chapter, you will be able to:

  • ✅ Understand the concept and technical pathways of Agentic-RL
  • ✅ Master the implementation of SFT + LoRA supervised fine-tuning
  • ✅ Deeply understand the principles of PPO, DPO, and GRPO reinforcement learning algorithms
  • ✅ Learn about DP / TP / PP / SP / ZeRO distributed training methods
  • ✅ Complete a full SFT + GRPO training pipeline
  • ✅ Understand intra-group sampling, relative advantage, and KL constraints in GRPO/RLVR through minimal runnable demos
  • ✅ Master data construction methods tailored for Agent fine-tuning (tool call trajectories, Gorilla/ToolBench datasets)
  • ✅ Track the latest research developments in the GRPO family, Agent RL, process rewards, and training efficiency optimization from 2025–2026

🔗 Learning Path

Prerequisites: Chapter 2: Large Language Model Foundations Recommended but not required: Chapter 5: Planning & Reasoning, Appendix E: KL Divergence Explained

Recommended Next Steps:


Chapter Overview

In previous chapters, we built Agents using prompts + tool calling — all of the Agent's capabilities came from the base model's pre-training knowledge combined with carefully designed prompts. This approach is simple and flexible, but it has a fundamental bottleneck:

The upper bound of an Agent's capability = the upper bound of the base model's general capability.

Agentic-RL (Agentic Reinforcement Learning) provides an alternative path: through reinforcement learning training, the model autonomously learns the optimal policy for completing Agent tasks. Works such as DeepSeek-R1 [1] and DeepSWE [2] have demonstrated that RL-trained models can exhibit reasoning strategies that never appeared in the training data, significantly outperforming pure prompt-based approaches in reasoning and tool-use capabilities.

📚 Chapter Structure

SectionContentKey Takeaways
10.1What is Agentic-RLUnderstand the essential differences between Agentic-RL and traditional post-training, master the MDP framework modeling approach
10.2SFT + LoRA Basic TrainingMaster the formal principles of supervised fine-tuning and LoRA parameter-efficient training methods
10.2bDistributed Training FundamentalsDP / TP / PP / SP / ZeRO distributed training methods
10.3PPO: Proximal Policy OptimizationStarting from policy gradients, systematically understand importance sampling, advantage functions, GAE, and the Clip mechanism
10.4DPO: Direct Preference OptimizationMaster the complete mathematical derivation from RLHF to DPO, understand the idea of implicit rewards
10.5GRPO/GSPO + Reward Function DesignUnderstand the principle of intra-group comparison replacing the Critic, along with multi-dimensional reward function design and reward hacking defense
10.6Hands-On: Complete Training PipelineFirst run through the core mechanisms with a minimal GRPO/RLVR demo, then complete the full training from data preparation to model deployment based on GSM8K
10.7Latest Research Developments (2025–2026)Survey cutting-edge works including DeepSeek-R1, DAPO, VAPO, SAR, T-STAR, UCPO, DelTA, SRPO, GROW, and grasp the technical landscape of the GRPO family and Agent RL
10.8Fine-Tuning Tailored for AgentsTool call data construction, three major open-source datasets (Gorilla/ToolBench/AgentInstruct), Agent SFT-specific configurations

📌 Extended Reading: The original 10.9 Agentic Data Flywheel, original 10.10/10.11 Self-Evolution Agent, and self-evolution frontier content have been independently expanded into Chapter 11: Self-Evolution and Continual Learning. It is recommended to continue reading after completing this chapter.


References

[1] DEEPSEEK AI. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning[R]. arXiv preprint arXiv:2501.12948, 2025.

[2] DEEPSEEK AI. DeepSWE: An open agentic SWE model that matches the performance of closed-source models[R]. 2025.