← Back to Research
arxiv
AI Agents February 25, 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 (core single-agent capabilities in stable environments), self-evolving agentic reasoning (agents refining capabilities through feedback and memory), and collective multi-agent reasoning (collaborative coordination and knowledge sharing). We further distinguish in-context reasoning from post-training reasoning, review frameworks across real-world applications including science, robotics, healthcare, and mathematics, and outline open challenges including personalization, long-horizon interaction, world modeling, scalable multi-agent training, and governance.
agentic reasoningLLM agentsmulti-agent systemsplanningreinforcement learning
Analysis

Summary

This comprehensive survey from a 29-author team (January 2026) provides the most up-to-date unified framework for understanding agentic reasoning in large language models. The authors organize the landscape along three layered dimensions: foundational single-agent reasoning (planning, tool use, search), self-evolving reasoning (memory, feedback loops, adaptation), and collective multi-agent reasoning (coordination, knowledge sharing). Crucially, they draw a sharp distinction between in-context reasoning — which scales test-time compute through structured orchestration — and post-training reasoning, which bakes better behaviors in via RL and SFT. This framing is more principled than most prior surveys.

Why It Matters

The field has been moving fast enough that no single prior survey covers the full arc from chain-of-thought prompting to autonomous multi-agent systems. This paper attempts to do exactly that, serving as a unified roadmap from "thought" to "action." The timing is significant: as of early 2026, agentic systems are no longer research prototypes — they are being deployed in healthcare, scientific research, and autonomous code generation. A rigorous taxonomy at this moment helps practitioners and researchers orient themselves in an increasingly fragmented landscape.

Implications

The survey's identification of open challenges is where it earns its keep. Long-horizon interaction remains deeply unsolved — current agents lose coherence over extended task horizons, and this is one of the primary limiters for real-world deployment. World modeling is another: agents that can only reason about what they've seen in training struggle when environments shift. The governance dimension is equally important; as collective multi-agent systems become more capable, questions of alignment, accountability, and emergent misalignment in agent networks become pressing. This paper provides researchers with the vocabulary and framework to address these problems systematically.

Critical Assessment

With 29 authors, surveys of this scope risk becoming reference lists rather than analytical works. The real test is whether the three-dimensional framework (foundational / self-evolving / collective) offers genuine predictive or organizational power, or whether it simply relabels existing categories. The in-context vs. post-training distinction is genuinely useful and underexplored in prior work. Overall this is a high-value orientation paper for anyone entering or repositioning within the agentic AI space in 2026.

Key Findings
1
Three-layer taxonomy: foundational, self-evolving, and collective multi-agent reasoning
2
Critical distinction between in-context (test-time) and post-training reasoning optimization
3
Long-horizon interaction, world modeling, and multi-agent governance identified as the primary unsolved challenges
Impact Score
8/10
Expert Commentary
The three-dimensional framework is the most coherent taxonomy of agentic reasoning I've seen — the in-context vs. post-training distinction alone makes this worth reading for anyone building agent systems in 2026.
D
Dong Tran
AI Researcher. Deep-dive analysis of frontier AI research papers.