Agentic Reasoning for Large Language Models
by Tianxin Wei et al. (arXiv:2601.12538)
The map of how LLMs became agents that plan, act and improve themselves.
Overview
A survey submitted to arXiv on 18 January 2026 (arXiv:2601.12538) with 29 authors led by Tianxin Wei, mapping how the field moved from LLMs reasoning inside a closed prompt to LLMs treated as autonomous agents that plan, act and learn through continual interaction. Its organising claim is that strong closed-world reasoning does not transfer to open-ended, dynamic environments, and it structures the literature along three complementary dimensions. Foundational agentic reasoning covers the single-agent capabilities, planning, tool-use optimisation and agentic search, in stable environments. Self-evolving agentic reasoning covers how agents refine themselves through feedback mechanisms, agentic memory, and evolution of their underlying capabilities. Collective multi-agent reasoning covers collaboration and division of labour plus shared multi-agent memory and evolution. Cutting across all three, the survey separates in-context reasoning, described as test-time scaling through structured orchestration and adaptive workflows, from post-training reasoning, described as behaviour optimisation via supervised fine-tuning and reinforcement learning; that distinction maps directly onto whether a practitioner can act on a technique with prompting alone or has to train. Planning subdivides further into workflow design, tree search and algorithm simulation, process formalisation, decomposition and external aid; tool use into interleaved reasoning, SFT bootstrapping, RL mastery and orchestration. Application chapters cover math exploration and coding agents, scientific discovery, embodied robotics, healthcare, and autonomous web research, each read through the same three lenses, followed by benchmarks and open challenges including personalisation and governance. A companion GitHub list, Awesome-Agentic-Reasoning, carries roughly 1.3k stars and keeps the catalogued papers current.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~3-5 hours to read; reference thereafter
- Provider
- Tianxin Wei et al. (arXiv:2601.12538)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓A three-dimension taxonomy: foundational, self-evolving and collective agentic reasoning
- ✓Why strong closed-world reasoning fails to transfer to open dynamic environments
- ✓The practical split between in-context reasoning and post-training reasoning approaches
- ✓How planning research divides into workflow design, tree search and decomposition
- ✓How tool-use training progresses from SFT bootstrapping to reinforcement-learning mastery
- ✓Which benchmarks the field uses to measure agentic reasoning capability
- ✓How the same techniques recur in science, robotics, healthcare and web-research agents
Highlights
- •29 authors from across the agent-reasoning research community, submitted January 2026, so it is current with the post-training and RL wave
- •Companion GitHub repository, Awesome-Agentic-Reasoning, with roughly 1.3k stars and a maintained paper list
- •The in-context versus post-training axis is the practically useful one: it tells you what you can act on with prompting alone
- •Application chapters are read through the same taxonomy, so techniques stay comparable across domains
- •Free preprint with no paywall, cross-listed under cs.AI and cs.CL
Who It’s For
Best For
- ✓Researchers orienting themselves in the agentic reasoning literature
- ✓Engineers who want the map before picking a planning or tool-use technique
- ✓Anyone deciding whether a problem needs prompting, fine-tuning or reinforcement learning
Prerequisites
- •Comfortable reading ML research papers and their notation
- •Working knowledge of LLM agents, tool use and reinforcement-learning basics
FAQ
What is Agentic Reasoning for Large Language Models?
A January 2026 survey with 29 authors organising agentic reasoning research along three dimensions: foundational single-agent capability, self-evolving agents, and collective multi-agent reasoning. It separates in-context techniques from post-training ones, covers applications across science, robotics, healthcare and web research, and ships with a maintained companion paper list on GitHub.
Is Agentic Reasoning for Large Language Models free?
Agentic Reasoning for Large Language Models is free to access.
What level is Agentic Reasoning for Large Language Models for?
Agentic Reasoning for Large Language Models is aimed at a advanced audience. Recommended background: Comfortable reading ML research papers and their notation, Working knowledge of LLM agents, tool use and reinforcement-learning basics.
How long does Agentic Reasoning for Large Language Models take?
Expect roughly ~3-5 hours to read; reference thereafter. Most learners work through it at their own pace.
What will I learn from Agentic Reasoning for Large Language Models?
You'll learn: A three-dimension taxonomy: foundational, self-evolving and collective agentic reasoning; Why strong closed-world reasoning fails to transfer to open dynamic environments; The practical split between in-context reasoning and post-training reasoning approaches; How planning research divides into workflow design, tree search and decomposition; How tool-use training progresses from SFT bootstrapping to reinforcement-learning mastery; Which benchmarks the field uses to measure agentic reasoning capability; How the same techniques recur in science, robotics, healthcare and web-research agents.
Topics
Sources
This page was written from 2 sources, 1 on domains other than arxiv.org.