arxiv
AI Agents
March 4, 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 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. This survey organizes agentic reasoning along three complementary dimensions: foundational agentic reasoning establishing core single-agent capabilities including planning, tool use, and search; self-evolving agentic reasoning studying how agents refine capabilities through feedback, memory, and adaptation; and collective multi-agent reasoning extending intelligence to collaborative settings. The survey further distinguishes in-context reasoning from post-training reasoning via reinforcement learning and supervised fine-tuning, and reviews frameworks across science, robotics, healthcare, autonomous research, and mathematics.
agentic reasoningLLM agentsmulti-agent systemsplanningtool usereinforcement learningsurvey
Analysis
Summary
This sweeping 29-author survey from UIUC (Wei et al., January 2026) attempts something ambitious: mapping the entire landscape of how LLMs reason when deployed as autonomous agents rather than static question-answerers. The key organizational insight is a three-layer taxonomy — foundational agentic reasoning (single-agent planning, tool use, search), self-evolving reasoning (feedback loops, memory, adaptation), and collective multi-agent reasoning (coordination, knowledge sharing, shared goals). This is the first survey I've seen that cleanly separates in-context reasoning (scaling at test time through orchestration) from post-training reasoning (optimizing via RL and SFT), which is a genuinely useful distinction for the field.
Why It Matters
We are in the middle of a Cambrian explosion of agentic frameworks — ReAct, Reflexion, AutoGPT, CrewAI, LangGraph, and dozens more — and the field desperately needs taxonomic clarity. This paper provides it. The three-layer decomposition isn't just academic bookkeeping; it maps directly onto real engineering decisions. When you're building an agent system, you need to decide: is this a single-agent planning problem, a self-improvement loop, or a multi-agent coordination challenge? The paper gives you the vocabulary and literature pointers to make that decision intelligently. The coverage of real-world applications (science, robotics, healthcare, autonomous research) is particularly valuable because it grounds the theoretical taxonomy in actual deployment contexts.
Implications
The open challenges section is where the real value lies for practitioners. The paper identifies personalization, long-horizon interaction, world modeling, scalable multi-agent training, and governance as the five frontier problems. I'd argue the governance piece is the most underdeveloped — we're deploying agent systems faster than we're building the oversight infrastructure to manage them. The distinction between in-context and post-training reasoning also has immediate practical implications: most production agent systems today rely entirely on in-context reasoning (prompt engineering, tool orchestration), but the post-training approaches (RLHF for agent behaviors, SFT on successful trajectories) are where the next performance leap will come from. This survey will be the standard reference for anyone entering the agentic AI space in 2026.
Key Findings
1
Three-layer taxonomy of agentic reasoning: foundational (planning, tool use, search), self-evolving (feedback, memory, adaptation), and collective multi-agent (coordination, knowledge sharing)
2
Critical distinction between in-context reasoning (test-time orchestration scaling) and post-training reasoning (RL/SFT optimization) as complementary paradigms
3
Five identified frontier challenges: personalization, long-horizon interaction, world modeling, scalable multi-agent training, and governance for real-world deployment
Expert Commentary
The most comprehensive agentic reasoning taxonomy published to date — essential reading for anyone building or researching LLM agent systems in 2026. The three-layer framework will become standard vocabulary.
D
Dong Tran
AI Researcher. Deep-dive analysis of frontier AI research papers.