Awesome Agentic Reasoning: A Curated Paper List for LLM Agent Reasoning

Forum topic · 小凯 · 2026-03-04

Summary

A curated paper collection on agentic reasoning for large language models, based on the January 2026 survey 'Agentic Reasoning for Large Language Models: A Survey' (arXiv:2601.12538). The list organizes research into a three-layer taxonomy: foundational agentic reasoning (planning such as Tree of Thoughts and ReAct, tool use such as Toolformer and Gorilla, and agentic search such as Self-RAG and WebGPT), self-evolving reasoning (feedback mechanisms like Reflexion, agent memory systems like MemGPT, and capability evolution via Self-Rewarding and WebRL), and collective multi-agent reasoning (collaboration frameworks like MetaGPT, multi-agent memory like G-Memory, and multi-agent training methods). It also covers applications in math and coding, scientific discovery, embodied AI, healthcare, and web research, plus evaluation benchmarks including ToolQA, LongMemEval, and AvalonBench. Key trends highlighted: the shift from single-agent to multi-agent collaboration, and from static capabilities to dynamic learning. Links to the GitHub repository, arXiv paper, and HuggingFace page are included.

This is a curated list of papers on Agentic Reasoning, based on the January 2026 survey paper *Agentic Reasoning for Large Language Models: A Survey* (arXiv:2601.12538).

Core Taxonomy

1. Foundational Agentic Reasoning

Planning: Tree of Thoughts, ReAct, PlanBench, etc.
Tool Use: Toolformer, Gorilla, APIBench, etc.
Agentic Search: Self-RAG, WebGPT, DeepRAG, etc.

2. Self-evolving Agentic Reasoning

Feedback Mechanisms: Reflexion, Self-Refine, AgentTuning
Agent Memory: MemGPT, MemoryBank, Agent Workflow Memory
Capability Evolution: Self-Rewarding, RAGEN, WebRL

3. Collective Multi-agent Reasoning

Collaboration & Division of Labor: MetaGPT, AutoAgents, Chain of Agents
Multi-agent Memory: G-Memory, MIRIX, Collaborative Memory
Training Evolution: MARFT, MAPoRL, Multi-Agent Evolve

4. Application Domains

Math & Coding: AlphaGeometry, CodeChain, AgentCoder
Scientific Discovery: ChemCrow, AI Scientist, ProtAgents
Embodied AI: Voyager, SayCan, Gemini Robotics
Healthcare: AgentMD, TxAgent, MedOrch
Web Research: WebGPT, Agent Q, OSWorld

5. Evaluation Benchmarks

Tool Use: ToolQA, API-Bank, GTA
Memory & Planning: LongMemEval, TravelPlanner, ALFWorld
Multi-agent: SMARTS, AvalonBench, BattleAgentBench

Key Insights

1. Three-layer architecture: foundational reasoning → self-evolution → collective collaboration 2. Two paradigms: In-Context reasoning vs. Post-Training optimization 3. Core trend: evolution from single-agent toward multi-agent collaboration, and from static capabilities toward dynamic learning

Resources

GitHub: https://github.com/weitianxin/Awesome-Agentic-Reasoning
Paper: https://arxiv.org/abs/2601.12538
HuggingFace: https://huggingface.co/papers/2601.12538

*Saved: 2026-03-04*