← Back to Research
arxiv
AI Agents February 26, 2026 Impact: 8/10

Agentic Reasoning for Large Language Models

By Tianxin Wei, Ting-Wei Li, Zhining Liu, Xuying Ning, Ze Yang, Jiaru Zou, Zhichen Zeng, Ruizhong Qiu, Xiao Lin, Dongqi Fu, Zihao Li, Mengting Ai, Duo Zhou, Wenxuan Bao, Yunzhe Li, Gaotang Li, Cheng Qian, Yu Wang, Xiangru Tang, Yin Xiao, Liri Fang, Hui Liu, Xianfeng Tang, Yuji Zhang, Chi Wang, Jiaxuan You, Heng Ji, Hanghang Tong, Jingrui He
Reasoning is a fundamental cognitive process underlying inference, problem-solving, and decision-making. While large language models (LLMs) demonstrate strong reasoning capabilities in closed-world settings, they struggle in open-ended and dynamic environments. Agentic reasoning marks a paradigm shift by reframing LLMs as autonomous agents that plan, act, and learn through continual interaction. In this survey, we organize agentic reasoning along three complementary dimensions: foundational agentic reasoning, self-evolving agentic reasoning, and collective multi-agent reasoning. We distinguish in-context reasoning from post-training reasoning, and review representative frameworks across science, robotics, healthcare, autonomous research, and mathematics.
agentic reasoningLLM agentsmulti-agent systemsplanningreinforcement learning
Analysis

Summary

This January 2026 survey from a 29-author team spanning UIUC, Stanford, Yale, and Microsoft Research offers the most comprehensive taxonomy of agentic reasoning to date. The authors organize the field along three nested layers: foundational single-agent capabilities (planning, tool use, search), self-evolving agents that refine behavior through feedback and memory, and collective multi-agent systems with coordination and shared goals. Crucially, they draw a clean line between in-context reasoning — scaling test-time interaction through structured orchestration — and post-training reasoning, which bakes behavior in via RL and SFT.

Why It Matters

The survey arrives at an inflection point. The field has been producing agentic frameworks faster than it can evaluate them, and the lack of a unified conceptual vocabulary has made it difficult to compare approaches or identify genuine progress. By building a three-dimensional roadmap, Wei et al. give researchers and practitioners a shared coordinate system. The distinction between in-context and post-training reasoning is especially valuable — it clarifies why the same underlying model can behave so differently depending on whether intelligence is scaffolded at runtime versus trained in.

Implications

For the near term, this survey highlights that the hardest open problems in agentic AI are not about raw capability but about sustained coherence: long-horizon interaction, personalization, and world modeling. The authors flag that most benchmarks still evaluate agents on short, closed tasks — a significant gap given that real deployment involves open-ended, multi-step environments. The governance section is notably substantive, reflecting growing field-wide awareness that autonomous agents operating in high-stakes domains (healthcare, robotics, research) require safety frameworks that don't yet exist.

Research Frontier

The survey's coverage of scalable multi-agent training stands out as the most underexplored frontier. While single-agent RL has matured significantly (DeepSeek-R1, o1, etc.), training agents that reliably coordinate at scale remains unsolved. The authors' project page promises an evolving resource as the field moves fast — worth tracking as a living index of what's current in agentic reasoning research.

Key Findings
1
Three-layer taxonomy: foundational, self-evolving, and collective multi-agent reasoning
2
Key distinction between in-context reasoning (test-time scaling) and post-training reasoning (RL/SFT)
3
Most benchmarks still evaluate on short closed tasks — major gap for real-world deployment
4
Long-horizon interaction and scalable multi-agent training identified as the hardest open problems
5
Governance frameworks for autonomous agents in high-stakes domains are critically underdeveloped
Impact Score
8/10
Expert Commentary
The most useful contribution here is the vocabulary, not the novelty — agentic AI desperately needs a shared conceptual framework, and this survey delivers one that should hold up as the field evolves.
D
Dong Tran
AI Researcher. Deep-dive analysis of frontier AI research papers.