awesome-agentic-reasoning

# Awesome Agentic Reasoning > Skill by [ara.so](https://ara.so) — AI Agent Skills collection. This skill provides expertise in navigating and utilizing the **Awesome Agentic Reasoning** repository — a comprehensive, curated collection of research papers and resources on agentic reasoning for Large Language Models (LLMs). The repository is based on the survey paper "[Agentic Reasoning for Large Language Models: A Survey](https://arxiv.org/abs/2601.12538)" and organizes cutting-edge research into foundational reasoning, self-evolving systems, and multi-agent collaboration. ## What This Repository Provides The Awesome Agentic Reasoning repository offers: - **Categorized Research Papers**: Organized by thematic areas including planning, tool use, search, self-evolution, multi-agent systems, and real-world applications - **Benchmarks**: Comprehensive lists of evaluation frameworks for agentic reasoning capabilities - **Three-Layer Framework**: - **Foundational Reasoning**: Core single-agent abilities (planning, tool-use, search) - **Self-Evolving Reasoning**: Adaptation through feedback, memory, and learning - **Collective Reasoning**: Multi-agent coordination and collaborative intelligence - **Application Domains**: Math/coding agents, scientific discovery, embodied agents, healthcare, web exploration - **Survey Materials**: Slides and the comprehensive survey paper ## Repository Structure ``` Awesome-Agentic-Reasoning/ ├── README.md # Main curated list ├── CONTRIBUTING.md # Contribution guidelines ├── materials/ # Survey slides and materials │ └── Agentic Reasoning Survey Talk.pdf └── figs/ # Framework diagrams ├── overview.png └── planning.png ``` ## Navigating the Repository ### Main Categories The repository organizes papers into three primary layers: #### 1. Foundational Agentic Reasoning **Planning Reasoning**: - In-context Planning (workflow design, tree search) - Post-training Planning (supervised fine-tuning, reinforcement learning) **Tool-Use Optimization**: - In-context Tool-Use (API orchestration, workflow design) - Post-training Tool-Use (supervised learning, RL fine-tuning) **Agentic Search**: - In-context Search (web navigation, knowledge retrieval) - Post-training Search (RL optimization) #### 2. Self-Evolving Agentic Reasoning - **Agentic Feedback Mechanisms**: Self-reflection, critique, and iterative refinement - **Agentic Memory**: Short-term and long-term memory systems - **Evolving Foundational Capabilities**: Continuous improvement of planning, tool-use, and search #### 3. Collective Multi-Agent Reasoning - **Role Taxonomy**: Debate, collaboration, hierarchical structures - **Collaboration Patterns**: Division of labor, coordination strategies - **Multi-Agent Memory and Evolution**: Shared knowledge, collective learning ### Applications The repository covers real-world applications: - 💻 **Math Exploration & Coding Agents** - 🔬 **Scientific Discovery Agents** - 🤖 **Embodied Agents** - 🏥 **Healthcare & Medicine Agents** - 🌐 **Autonomous Web Exploration & Research Agents** ### Benchmarks Organized by: - **Core Mechanisms**: Tool Use, Search, Memory & Planning, Multi-Agent Systems - **Application Domains**: Embodied, Scientific Discovery, Medical, Web, General Tool-Use ## Usage Patterns ### Finding Papers on Specific Topics **Example 1: Finding Planning Papers** Navigate to the Planning Reasoning section to find papers on: - Workflow design approaches (ReAct, ReWOO, Plan-and-Solve) - Tree search methods (Tree of Thoughts, MCTS-based approaches) - Post-training planning optimization **Example 2: Multi-Agent System Research** The Collective Multi-Agent Reasoning section includes: - Role specialization papers - Collaboration frameworks - Multi-agent memory systems ### Exploring Application Domains **Example: Embodied Agent Research** 1. Check the **Applications > Embodied Agents** section 2. Cross-reference with **Benchmarks > Embodied Agents** for evaluation frameworks 3. Review foundational papers on planning and tool-use that apply to embodied settings ### Finding Benchmarks **Example: Evaluating Tool-Use Capabilities** ```markdown ## Tool Use Benchmarks Navigate to: Benchmarks > Core Mechanisms > Tool Use Key benchmarks include: - API-Bank: API selection and execution - ToolBench: Multi-tool orchestration - T-Eval: Tool learning evaluation ``` ## Contributing to the Repository ### Adding New Papers Create a pull request with papers organized by category: ```markdown | [Paper Title](https://arxiv.org/abs/XXXX.XXXXX) | Conference/Year | ``` **Guidelines**: - Place papers in the appropriate thematic section - Follow the existing table format - Include the full arXiv link or conference proceedings URL - Add the publication year or venue ### Suggesting Resources Open an issue to suggest: - New paper categories - Additional benchmarks - Application domains not yet covered - Survey materials or tutorials **Contact**: - Email: twei10@illinois.edu, twli@illinois.edu, liu326@illinois.edu - GitHub Issues: For suggestions and discussions ## Key Research Paradigms ### In-Context Reasoning vs. Post-Training Reasoning The repository distinguishes between two optimization approaches: **In-Context Reasoning**: - Test-time scaling through structured orchestration - Adaptive workflows without parameter updates - Examples: ReAct, Tree of Thoughts, Chain-of-Thought prompting **Post-Training Reasoning**: - Behavior optimization via RL and supervised fine-tuning - Parameter updates to internalize reasoning strategies - Examples: RLHF for tool-use, Q-learning for planning ### Environmental Dynamics Papers are organized by the environmental setting: - **Static environments**: Fixed tool sets, deterministic outcomes - **Dynamic environments**: Feedback loops, adaptation requirements - **Multi-agent environments**: Coordination, communication, emergent behavior ## Working with Survey Materials ### Accessing the Survey Paper The foundational survey is available at: - arXiv: https://arxiv.org/abs/2601.12538 - HuggingFace Papers: https://huggingface.co/papers/2601.12538 ### Using the Slides Presentation materials are in `materials/Agentic Reasoning Survey Talk.pdf`: - Framework overview - Key insights from each reasoning layer - Application case studies - Future research directions ## Common Patterns ### Building a Research Bibliography **Pattern: Comprehensive Literature Review** ```python # Pseudo-code for extracting papers by category categories = [ "Planning Reasoning", "Tool-Use Optimization", "Agentic Search", "Multi-Agent Systems" ] papers_by_category = {} for category in categories: # Navigate to README section papers = extract_papers_from_section(category) papers_by_category[category] = papers # Generate BibTeX or reading list ``` ### Tracking New Research **Pattern: Monitoring Updates** The repository is actively maintained. To stay current: 1. Watch the repository for updates 2. Check the News section in README for announcements 3. Review recent commits for newly added papers 4. Subscribe to GitHub notifications ### Cross-Referencing Applications and Benchmarks **Pattern: Application-Specific Research** For a specific application domain: ```markdown 1. Identify application section (e.g., "Healthcare & Medicine Agents") 2. Review papers in that section 3. Navigate to corresponding benchmark section 4. Check foundational techniques used (planning, tool-use, etc.) 5. Trace back to foundational reasoning sections for core methods ``` ## Citation When using this repository in research or projects: ```bibtex @article{wei2026agentic, title={Agentic Reasoning for Large Language Models}, author={Wei, Tianxin and Li, Ting-Wei and Liu, Zhining and Ning, Xuying and Yang, Ze and Zou, Jiaru and Zeng, Zhichen and Qiu, Ruizhong and Lin, Xiao and Fu, Dongqi and others}, journal={arXiv preprint arXiv:2601.12538}, year={2026} } ``` ## Integration with Development Workflows ### For Researchers **Literature Review Workflow**: 1. Clone the repository for offline access 2. Use the categorized structure to identify relevant papers 3. Cross-reference applications with foundational techniques 4. Export citations for reference management tools ### For Practitioners **Implementation Workflow**: 1. Identify your application domain (e.g., web agents, coding) 2. Review application-specific papers and benchmarks 3. Trace foundational techniques (planning, tool-use) 4. Reference implementation papers for code patterns 5. Evaluate using suggested benchmarks ### For Tool Builders **Benchmark Selection**: 1. Determine core capability (planning, tool-use, search) 2. Navigate to corresponding benchmark section 3. Review evaluation frameworks and metrics 4. Compare agent performance across standard benchmarks ## Best Practices ### Exploring New Topics 1. **Start with the Overview**: Read the survey paper introduction and framework diagram 2. **Navigate by Layer**: Begin with foundational reasoning before advanced topics 3. **Cross-Reference**: Link application papers back to foundational techniques 4. **Check Benchmarks**: Understand evaluation standards for each capability ### Contributing Quality Additions 1. **Verify Relevance**: Ensure papers fit the agentic reasoning scope 2. **Check Duplicates**: Search existing entries before adding 3. **Provide Context**: Include venue/year information 4. **Follow Format**: Maintain consistent table structure ### Staying Current 1. **Monitor Commits**: The repository updates regularly with new papers 2. **Check News Section**: Major updates announced at the top of README 3. **Watch Discussions**: GitHub issues may highlight emerging trends 4. **Follow Survey Updates**: Authors plan continued improvements ## Troubleshooting ### Finding Specific Papers **Issue**: Can't locate a specific paper **Solution**: - Use browser search (Ctrl+F / Cmd+F) on the README - Check multiple related sections (papers may fit several categories) - Review the benchmarks section for evaluation-focused papers - Check recent commits if it's a new publication ### Understanding Categories **Issue**: Unclear which section contains relevant papers **Solution**: - Refer to the framework overview diagram - Read the category descriptions in the survey paper - Cross-reference with similar known papers - Check application sections if domain-specific ### Accessing Papers **Issue**: Links not working or papers behind paywalls **Solution**: - Most papers link to arXiv versions (open access) - For conference papers, search on Google Scholar - Check author websites for preprints - Use institutional access for published versions ## Related Resources - **Survey Paper**: https://arxiv.org/abs/2601.12538 - **Presentation Slides**: `materials/Agentic Reasoning Survey Talk.pdf` - **HuggingFace**: https://huggingface.co/papers/2601.12538 - **License**: MIT (open for use and contribution) ## Quick Reference | Category | Key Papers | Benchmarks | |----------|-----------|------------| | Planning | Tree of Thoughts, ReAct, Plan-and-Solve | PlanBench, BlocksWorld | | Tool-Use | Gorilla, ToolLLM, HuggingGPT | API-Bank, ToolBench | | Search | WebGPT, Agent-E, Mind2Web | WebArena, GAIA | | Multi-Agent | ChatDev, AgentVerse, MetaGPT | MAgIC, AgentBench | | Embodied | LM-Nav, PERIA, RT-1 | CALVIN, MetaWorld | | Scientific | FunSearch, AI Scientist | ScienceBench | This skill enables AI coding agents to effectively navigate and utilize the Awesome Agentic Reasoning repository, helping developers access cutting-edge research on LLM-based agents, understand agentic reasoning frameworks, and apply state-of-the-art techniques to their projects.