DOCSEARCHbyAlgolia
Submit Docs

Awesome Agentic Reasoning

repository·main·Indexed 23 days ago

https://github.com/weitianxin/awesome-agentic-reasoning

A curated collection of research papers and resources regarding Agentic Reasoning for Large Language Models. Based on the survey paper 'Agentic Reasoning for Large Language Models' (arXiv:2601.12538), the repository organizes research into foundational reasoning (planning, tool-use, search), self-evolving reasoning (feedback, memory), and collective reasoning (multi-agent coordination), covering both in-context and post-training optimization paradigms.

Tokens
10.1K
Snippets
6
Records
29
Agent score
30%

What's inside awesome-agentic-reasoning

  1. Overview of Awesome Agentic Reasoning Papers

    main

    This repository is a curated collection of research papers focused on Agentic Reasoning, which integrates reasoning with action. It organizes research into thematic areas such as planning, tool use, search, self-evolution (via memory and feedback), multi-agent systems, and real-world applications/benchmarks.

    The collection is based on the survey paper: Agentic Reasoning for Large Language Models.

  2. Research papers on Post-training Collaboration

    main

    This section covers methods for optimizing multi-agent systems after the initial model training, focusing on three optimization strategies:

    1. Multi-agent Prompt Optimization: Techniques to improve agent performance by optimizing prompts and topologies (e.g., DSPy Assertions, Multi-agent Design).
    2. Graph-based Topology Generation: Research on architecting the communication structure between agents using graphs, including GNN-based design (G-Designer), graph diffusion (Graph Diffusion), and automated workflow generation (AFlow).
    3. Policy-based Topology Generation: Using Reinforcement Learning (RL) to learn optimal routing and coordination policies (e.g., MASRouter, xRouter, LLM Collaboration with Multi-Agent Reinforcement Learning).

  3. Research papers on Multi-Agent Memory and Evolution

    main

    This section explores how agents (single or multi-agent) evolve their capabilities and manage memory over time. Key research areas include:

    1. Evolutionary Patterns:
      • Intra-test-time Evolution: Improving performance during a single task execution via feedback (e.g., Reflexion, Self-Refine).
      • Inter-test-time Evolution: Improving capabilities across different tasks/sessions (e.g., STaR, RAGEN).
      • Multi-agent Evolution: Systems where multiple agents co-evolve or improve collectively (e.g., TextGrad, REMA).
    2. Memory Management: Research on hierarchical memory (G-Memory), structured contextual memory (Intrinsic Memory Agents), and decentralized/shared memory systems (Collaborative Memory, MIRIX).
    3. Training for Evolution: Methods to train agents specifically to evolve, such as Multi-Agent Reinforcement Fine-Tuning (MARFT) and Co-Evolving Multi-Agent Systems (CoMAS).

  4. Research Papers: Scientific Discovery & Multi-Agent Reasoning

    main

    This repository provides a curated list of research papers focused on agentic reasoning in scientific domains. Key research areas include:

    • Scientific Discovery: Using LLMs for chemical reasoning, materials design, and physical scientific discovery (e.g., ChemAgent, LLMatDesign).
    • Collective Multi-Agent Reasoning: Frameworks for multi-agent collaboration in scientific research, such as protein discovery (ProtAgents) and single-cell analysis (CellAgent).
    • Embodied Agents: Research on grounding language in robotic affordances, task planning, and vision-language-action models (e.g., SayPlan, EmbodiedGPT).
    • Self-evolving Agentic Reasoning: Agents capable of long-horizon tasks, memory-augmented planning, and lifelong learning (e.g., Voyager, Ella).
    • Multi-Agent Collaboration in Robotics: Cooperative plan optimization and theory-of-mind for heterogeneous multi-robot systems (e.g., CaPo, COHERENT).

  5. Explore Agentic Reasoning Research Papers

    main
    This repository provides a curated list of research papers focused on Agentic Reasoning. The papers are categorized into several key domains of agent development, including foundational reasoning, planning, web navigation, GUI automation, and scientific research automation. Developers can use this list to stay updated on the latest methodologies such as Reinforcement Learning (RL) for agents, multi-modal GUI grounding, and autonomous scientific discovery.

  6. Research papers on Collective Multi-agent Reasoning

    main

    This section categorizes research papers focused on how multiple agents collaborate to solve tasks. The research is divided into three main approaches to in-context collaboration:

    1. Manually Crafted Pipelines: Research on predefined workflows, such as hierarchical frameworks (e.g., AgentOrchestra), collaborative frameworks (e.g., MetaGPT), and specialized RAG workflows (e.g., Collab-RAG, MA-RAG).
    2. LLM-Driven Pipelines: Research on systems where LLMs dynamically manage the multi-agent process, including automatic agent generation (AutoAgents), generalist systems (Magentic-One), and routing mechanisms (AgentRouter, Talk to Right Specialists).
    3. Theory-of-Mind-Augmented Collaboration: Research exploring how agents use 'Theory of Mind' to understand and predict the mental states/intentions of other agents to improve coordination (e.g., BeliefNest, MindForge).

  7. Understand the Three Layers of Agentic Reasoning

    main

    Agentic reasoning is organized into three distinct layers based on environmental dynamics:

    1. Foundational Reasoning: Focuses on core single-agent abilities like planning, tool-use, and search within environments.
    2. Self-Evolving Reasoning: Focuses on adaptation through feedback, memory, and learning in dynamic settings.
    3. Collective Reasoning: Focuses on multi-agent coordination, role specialization, and collaborative intelligence.

    Additionally, reasoning paradigms are categorized by their optimization settings:

    • In-Context Reasoning: Test-time scaling through structured orchestration and adaptive workflows.
    • Post-Training Reasoning: Behavior optimization via Reinforcement Learning (RL) and supervised fine-tuning.

  8. Explore Agentic Memory Strategies

    main

    Agentic memory is crucial for long-horizon tasks and is categorized by how memory is stored and utilized:

    1. Flat Memory:
      • Factual Memory: Using retrieval-augmented generation (RAG) or long-term storage (e.g., LlamaIndex, MemGPT, MemoryBank).
      • Experience Memory: Storing workflows and test-time learning data (e.g., Agent Workflow Memory, ReasoningBank).
    2. Structured Memory: Organizing memory into graphs or hierarchical schemas (e.g., RepoGraph, Graph RAG, Mem0, Zep).
    3. Post-training Memory Control: Using reinforcement learning to teach agents how to manage and utilize their own memories (e.g., Memory-R1, MemRL, MemAgent).

  9. Explore Self-evolving Foundational Agentic Capabilities

    main

    Foundational capabilities can be evolved through self-improvement loops in several domains:

    1. Self-evolving Planning: Agents improve their ability to plan via self-rewarding, self-challenging, or differentiable text feedback (e.g., TextGrad, Self-rewarding language models).
    2. Self-evolving Tool-use: Agents learn to create their own specialized tools to solve complex problems (e.g., Large Language Models as Tool Makers, CREATOR).
    3. Self-evolving Search for Memory Retrieval: Improving how agents search through their own memory to find relevant context (e.g., Self-RAG, Mem0, Memory as Action).

  10. Explore Agentic Feedback Mechanisms

    main

    Agentic feedback mechanisms allow agents to self-evolve through different types of reinforcement. Research in this area is categorized into three main approaches:

    1. Reflective Feedback: Agents use verbal reinforcement or iterative refinement to improve their reasoning (e.g., Reflexion, Self-Refine, Tree of Thoughts).
    2. Parametric Adaptation: Improving the model's underlying parameters through techniques like fine-tuning or preference optimization (e.g., AgentTuning, Direct Preference Optimization (DPO), Constitutional AI).
    3. Validator-Driven Feedback: Using external validators, such as code execution or environment feedback, to guide the agent (e.g., CodeRL, LEVER, SWE-bench).

  11. Explore Scientific Discovery Agent Papers

    main

    This section provides a curated list of research papers focused on agents designed for scientific research and discovery. The papers are categorized into two research directions:

    1. Foundational Agentic Reasoning: Covers general-purpose scientific agents, tool-augmented reasoning (SciAgent), and domain-specific agents for chemistry (Chemcrow, CACTUS), biology (Biomni), and materials science (LLaMP, HoneyComb).
    2. Self-evolving Agentic Reasoning: (Note: The provided segment ends at the header for this section).

  12. Explore Math Exploration & Vibe Coding Agent Papers

    main

    This section provides a curated list of research papers focused on agents for mathematical reasoning and code generation (vibe coding). The papers are categorized into three research directions:

    1. Foundational Agentic Reasoning: Focuses on core capabilities like tool use (e.g., Toolformer), repository-level coding (CodePlan), and mathematical discovery.
    2. Self-evolving Agentic Reasoning: Focuses on agents that improve through feedback, such as self-debugging (Teaching Large Language Models to Self-Debug) and iterative refinement (Self-Refine).
    3. Collective Multi-agent Reasoning: Focuses on multi-agent frameworks for collaborative tasks like pair programming (AgentCoder) or securing code generation (AutoSafeCoder).