🧠 model-native-agentic-ai

<div align="center"> <img src="logo.jpg" alt="Logo" width="800"> <h1 align="center">Agentic AI: From Pipelines to Model-native</h1> <p align="center"> This repository contains a curated list of papers referenced in our survey: <br> <a href="https://arxiv.org/pdf/2510.16720v2"><strong>Beyond Pipelines: A Survey of the Paradigm Shift toward Model-Native Agentic AI</strong></a><br> We will continuously update this list with new, relevant papers. </p>

</div>

🙏 Citation

If you find our survey useful for your research, please consider citing our work:

@misc{sang2025pipelinessurveyparadigmshift,
      title={Beyond Pipelines: A Survey of the Paradigm Shift toward Model-Native Agentic AI}, 
      author={Jitao Sang and Jinlin Xiao and Jiarun Han and Jilin Chen and Xiaoyi Chen and Shuyu Wei and Yongjie Sun and Yuhang Wang},
      year={2025},
      eprint={2510.16720},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2510.16720}, 
}

🌟 Abstract

The rapid evolution of agentic AI marks a new phase in artificial intelligence, where Large Language Models (LLMs) no longer merely respond but act, reason, and adapt. This survey traces the paradigm shift in building agentic AI: from Pipeline-based systems, where planning, tool use, and memory are orchestrated by external logic, to the emerging Model-native paradigm, where these capabilities are internalized within the model’s parameters.

We first position Reinforcement Learning (RL) as the algorithmic engine enabling this paradigm shift. By reframing learning from imitating static data to outcome-driven exploration, RL underpins a unified solution of LLM + RL + Task across language, vision and embodied domains. Building on this, the survey systematically reviews how each capability—Planning, Tool use, and Memory—has evolved from externally scripted modules to end-to-end learned behaviors. Furthermore, it examines how this paradigm shift has reshaped major agent applications, specifically the Deep Research agent emphasizing long-horizon reasoning and the GUI agent emphasizing embodied interaction.

We conclude by discussing the continued internalization of agentic capabilities like Multi-agent collaboration and Reflection, alongside the evolving roles of the system and model layers in future agentic AI. Together, these developments outline a coherent trajectory toward model-native agentic AI as an integrated learning and interaction framework, marking the transition from constructing systems that apply intelligence to developing models that grow intelligence through experience.

📒 Table of Contents

• Core Capabilities: Planning
  • Pipeline-based Paradigm
  • Model-native Paradigm
• Core Capabilities: Tool Use
  • Pipeline-based Paradigm
  • Model-native Paradigm
• Core Capabilities: Memory
  • Short-Term Memory: Long Context
  • Short-Term Memory: Context Management
  • Long-Term Memory
• Applications
  • Deep Research Agent
  • GUI Agent
• Future Direction and Discussion: Emerging Model-native Capabilities
  • Multi-agent Collaboration
  • Reflection

📜 Papers

➤ 3  Core Capabilities: Planning

3.2  Pipeline-based Paradigm

| Short Name | Paper | Date | Code/Project Link | | :---------: | :---: | :--: | :---------------: | | STRIPS | Strips: A new approach to the application of theorem proving to problem solving | 1971 | - | | PDDL | PDDL - The Planning Domain Definition Language | 1998 | - | | LLM+P | LLM+P: Empowering Large Language Models with Optimal Planning Proficiency | 2023-04 | GitHub | | LLM+PDDL | Leveraging pre-trained large language models to construct and utilize world models for model-based task planning | 2023-05 | GitHub | | CoT | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models | 2022-01 | - | | ToT | Tree of Thoughts: Deliberate Problem Solving with Large Language Models | 2023-05 | GitHub | | RAP | Reasoning with Language Model is Planning with World Model | 2023-05 | - | | LLM+MCTS | Planning with MCTS: Enhancing Problem-Solving in Large Language Models | 2024-09 | - |

3.3  Model-native Paradigm

| Short Name | Paper | Date | Code/Project Link | | :---------: | :---: | :--: | :---------------: | | ReST-MCTS* | ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search | 2024-06 | GitHub | | Marco-o1 | Marco-o1: Towards Open Reasoning Models for Open-Ended Solutions | 2024-11 | GitHub | | HuatuoGPT-o1 | HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs | 2024-12 | GitHub | | Bespoke-Stratos | Bespoke-Stratos: The unreasonable effectiveness of reasoning distillation | 2024-12 | Dataset | | s1 | s1: Simple test-time scaling | 2025-01 | GitHub | | R1-Distill-SFT | Millions scale dataset distilled from R1-32b | 2025-01 | Dataset | | LIMO | LIMO: Less is More for Reasoning | 2025-02 | GitHub | | BOLT | BOLT: Bootstrap Long Chain-of-Thought in Language Models without Distillation | 2025-02 | - | | AStar | Boosting Multimodal Reasoning with Automated Structured Thinking | 2025-02 | - | | FastMCTS | FastMCTS: A Simple Sampling Strategy for Data Synthesis | 2025-02 | - | | OpenThoughts | OpenThoughts: Data Recipes for Reasoning Models | 2025-06 | GitHub | | OpenR1-Math-220k | Open R1: A fully open reproduction of DeepSeek-R1 | 2025-02 | GitHub | | SYNTHETIC-1 | SYNTHETIC-1: Two Million Collaboratively Generated Reasoning Traces from Deepseek-R1 | 2025-02 | Dataset | | WebSynthesis | WebSynthesis: World-Model-Guided MCTS for Efficient WebUI-Trajectory Synthesis | 2025-07 | GitHub | | Math-Shepherd | Math-Shepherd: Verify and Reinforce LLMs Step-by-step without Human Annotations | 2023-12 | Project | | ReFT | ReFT: Reasoning with Reinforced Fine-Tuning | 2024-01 | GitHub | | OmegaPRM | Improve Mathematical Reasoning in Language Models by Automated Process Supervision | 2024-06 | - | | OpenAI o1 | Learning to reason with LLMs | 2024-09 | - | | RLEF | [RLEF: Grounding Code LLMs in Execution Feedback wit