The Landscape of Agentic Reinforcement Learning for LLMs: A Survey
by arXiv
A comprehensive map of agentic reinforcement learning — how RL turns LLMs into autonomous decision-making agents.
Overview
The Landscape of Agentic Reinforcement Learning for LLMs: A Survey reframes reinforcement learning for large language models around agency — the shift from treating LLMs as passive sequence generators aligned with preference-based RL toward autonomous, decision-making agents operating in temporally extended, partially observable Markov decision processes (POMDPs). Synthesizing roughly 500 recent works, it organizes the field along two axes: a set of core agentic capabilities (planning, tool use, memory, reasoning, self-improvement, and perception) and the task domains where RL optimizes those capabilities. The survey also catalogs the open-source environments, benchmarks, and frameworks that support agentic RL research, making it a practical entry point and reference for anyone building or studying RL-trained LLM agents. First submitted September 2025 and revised through April 2026, it reflects a current view of the area.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- Long-form survey (~500 works reviewed)
- Provider
- arXiv
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How agentic RL differs from preference-based LLM alignment (POMDP vs. bandit framing)
- ✓A taxonomy of agentic capabilities RL targets: planning, tool use, memory, reasoning, self-improvement, perception
- ✓Which task domains agentic RL is applied to and how they map to those capabilities
- ✓The open-source environments, benchmarks, and frameworks available for agentic RL research
- ✓A structured reading map across ~500 recent papers in the field
Highlights
- •Synthesizes ~500 works into a single coherent taxonomy
- •Actively maintained — first submitted Sept 2025, revised through April 2026
- •Bridges RL theory and practical agent-building capabilities
Who It’s For
Best For
- ✓Researchers and advanced engineers working on LLM agents and RL
- ✓Practitioners choosing benchmarks, environments, or frameworks for agentic RL
- ✓Anyone needing a current, structured literature map of the field
Prerequisites
- •Solid understanding of reinforcement learning fundamentals
- •Familiarity with LLMs, agents, and tool use
FAQ
What is The Landscape of Agentic Reinforcement Learning for LLMs: A Survey?
A 2025 arXiv survey (revised April 2026) that maps the fast-moving field of agentic reinforcement learning for LLMs. It's for researchers and advanced engineers who want a structured, citation-rich overview of how RL trains LLMs to plan, use tools, remember, reason, and self-improve.
Is The Landscape of Agentic Reinforcement Learning for LLMs: A Survey free?
The Landscape of Agentic Reinforcement Learning for LLMs: A Survey is free to access.
What level is The Landscape of Agentic Reinforcement Learning for LLMs: A Survey for?
The Landscape of Agentic Reinforcement Learning for LLMs: A Survey is aimed at a advanced audience. Recommended background: Solid understanding of reinforcement learning fundamentals, Familiarity with LLMs, agents, and tool use.
How long does The Landscape of Agentic Reinforcement Learning for LLMs: A Survey take?
Expect roughly Long-form survey (~500 works reviewed). Most learners work through it at their own pace.
What will I learn from The Landscape of Agentic Reinforcement Learning for LLMs: A Survey?
You'll learn: How agentic RL differs from preference-based LLM alignment (POMDP vs. bandit framing); A taxonomy of agentic capabilities RL targets: planning, tool use, memory, reasoning, self-improvement, perception; Which task domains agentic RL is applied to and how they map to those capabilities; The open-source environments, benchmarks, and frameworks available for agentic RL research; A structured reading map across ~500 recent papers in the field.