Awesome-AgenticLLM-RL-Papers

This is the Official repo for the survey paper: The Landscape of Agentic Reinforcement Learning for LLMs: A Survey

ArXiv – https://arxiv.org/abs/2509.02547

HuggingFace – https://huggingface.co/papers/2509.02547

Citation

@article{
      zhang2026landscapeagenticreinforcementlearning,
      title={The Landscape of Agentic Reinforcement Learning for {LLM}s: A Survey},
      author={Guibin Zhang and Hejia Geng and Xiaohang Yu and Zhenfei Yin and Zaibin Zhang and Zelin Tan and Heng Zhou and Zhong-Zhi Li and Xiangyuan Xue and Yijiang Li and Yifan Zhou and Yang Chen and Chen Zhang and Yutao Fan and Zihu Wang and Songtao Huang and Francisco Piedrahita Velez and Yue Liao and Hongru WANG and Mengyue Yang and Heng Ji and Jun Wang and Shuicheng YAN and Philip Torr and LEI BAI},
      journal={Transactions on Machine Learning Research},
      issn={2835-8856},
      year={2026},
      url={https://openreview.net/forum?id=RY19y2RI1O},
      note={Survey Certification}
}

Sec2.7 Agentic RL: Algorithms

Clip corresponds to preventing the policy ratio from moving too far from 1 for ensuring stable updates.
KL penalty corresponds to penalizing the KL divergence between the learned policy and the reference policy for ensuring alignment.

Sec4.1 Task: Search & Research Agent

---

Sec4.2 Task: Code Agent

---

Sec4.3 Task: Mathematical Agent

---

Sec4.4 Task: GUI Agent

---

Sec4.5 Task: RL in Vision Agents

TO BE ADDED

---

Sec4.6 Task: RL in Embodied Agents

TO BE ADDED

---

Sec4.7 Task: RL in Multi-Agent Systems

“Dynamic” denotes whether the multi-agent system is task-dynamic, i.e., processes different task queries with different configurations (agent count, topologies, reasoning depth, prompts, etc).
“Train” denotes whether the method involves training the LLM backbone of agents.

Sec4.8. Task: Other Tasks

TO BE ADDED

Sec5.1 Environments

The agent capabilities are denoted by:
① Reasoning, ② Planning, ③ Tool Use, ④ Memory, ⑤ Collaboration, ⑥ Self-Improve.

Sec5.2 Frameworks