A Comprehensive Literature Survey on Large Language Model-Driven Operating Systems

Abstract

Recent progress in artificial intelligence has reshaped the design philosophy of modern operating systems, with Large Language Models (LLMs) emerging as a key enabling technology. Existing operating systems primarily rely on predefined graphical and command-line interactions, which limits their ability to adapt to user preferences, understand context, and respond to complex intent. Early natural language–based interfaces attempted to address these limitations; however, they lacked robust contextual reasoning and system-level intelligence required for dynamic computing environments. As workloads and user expectations continue to grow in complexity, there is an increasing need for operating systems capable of understanding and acting upon natural language instructions in a reliable and transparent manner. This paper examines operating system architectures that embed LLMs within core system components to facilitate intent-driven interaction and intelligent automation. The study surveys representative frameworks such as AIOS, PEROS, Compressor–Retriever architectures, and Herding LLaMaS, emphasizing how these designs translate linguistic input into executable operating system functions. Key capabilities explored include continuous context retention, dynamic allocation of computational resources, and semantic coordination between software services and underlying hardware. These design choices aim to enhance usability while reducing the cognitive burden placed on end users. The discussion further addresses essential system considerations, including interpretability of model-driven decisions, user-specific adaptation, protection of sensitive data, and scalability across diverse computing platforms. Emphasis is placed on explainable reasoning mechanisms that help users and developers understand system behavior, particularly in safety- and security-critical scenarios. By integrating adaptable intelligence with transparent control logic, LLM-enhanced operating systems can improve trust and operational efficiency.

Introduction

Modern operating systems (OS) act as the central coordination layer between users, software, and hardware. Traditional interfaces—graphical or command-line—require precise commands, limiting accessibility and adaptability. Large Language Models (LLMs) are transforming OS interaction by enabling natural language communication, allowing systems to infer user intent, translate high-level goals into system actions, and adapt dynamically to user context. LLM-driven OS architectures promise enhanced usability, personalization, efficiency, and resource optimization, but they raise challenges in transparency, reliability, and security.

Key Research Contributions in LLM-based Operating Systems:

1. AIOS (LLM Agent Operating Systems):

   • Introduces a structured, OS-inspired framework to manage LLM agents like conventional processes.

   • Features kernel-level control for resource allocation, execution isolation, task scheduling, and context management.

   • Agent requests are decomposed into modular operations with fair scheduling and memory management.

   • Achieves up to 2.1× throughput improvement, managing ~2,000 concurrent agents while maintaining accuracy and stability.

2. PEROS (Personalized Self-Adapting OS in the Cloud):

   • Focuses on personalization, autonomous adaptation, and privacy in cloud and multi-device environments.

   • Uses LLMs to interpret natural language user intent and dynamically configure kernel-level policies.

   • Microkernel-based design ensures modularity and fault isolation; cloud backends enable elastic scaling.

   • Integrates privacy-preserving data handling (“Databoxes”) to maintain user trust while delivering personalized functionality.

   • Demonstrates dynamic kernel adaptation and improved user experience under varying workloads.

3. Compressor–Retriever Architecture:

   • Addresses LLM context limitations by compressing and retrieving long-term system and user histories.

   • Converts historical interactions into compact embeddings stored in expandable memory repositories.

   • Sparse retrieval and attention mechanisms allow LLMs to reason over relevant context efficiently.

   • Experimental evaluation shows ~75% accuracy retention compared to full-context models, supporting scalable long-term reasoning.

4. Herding LLaMaS (LLMs as OS Modules):

   • LLMs function as OS components to manage heterogeneous hardware without manual configuration.

   • Natural language descriptions of device properties are embedded and used for memory allocation, task scheduling, and data placement.

   • Enables adaptive, automated handling of diverse devices across cloud, edge, and data-center infrastructures.

Overall Insight:
LLM-driven operating systems are evolving from conventional command-based models to intelligent, intent-aware platforms capable of self-adaptation, personalization, and efficient resource coordination. Key innovations include kernel-level management of LLM agents, privacy-preserving personalization, context-compression strategies for long-term reasoning, and LLM-based hardware management. These systems highlight the potential of integrating AI deeply into OS architectures, improving usability, scalability, and operational efficiency while raising important questions around transparency and security.

Conclusion

The AIOS framework shows that foundational operating system mechanisms—such as task scheduling, concurrency management, and memory handling—can be effectively reengineered to support large-scale execution of LLM-based agents. This work primarily advances system-level performance and scalability. In contrast, PEROS places emphasis on the user-facing dimension, demonstrating how adaptive kernels and natural language interfaces can deliver personalized experiences while enforcing strong privacy protections. The Compressor–Retriever approach addresses a fundamental limitation of LLM-based systems by enabling persistent, long-term context handling, thereby supporting reasoning beyond isolated interaction sessions. Meanwhile, Herding LLaMaS highlights the role of semantic understanding in hardware management, illustrating how LLMs can simplify integration and coordination across heterogeneous computing devices. A comparative analysis reveals that these approaches offer complementary contributions. AIOS excels in kernel-level control and concurrency but pays limited attention to personalization. Overall, existing research points toward a future in which operating systems are built around natural language interaction, adaptive reasoning, and interpretable decision processes. The insights from these studies underscore the importance of developing operating systems that are not only technically robust but also transparent and user-centric. Advancing explainability, efficiency, and trustworthiness will be critical in realizing LLM-driven operating systems that can reliably interpret user intent, manage resources intelligently, and communicate system behavior in ways that users can readily understand

References

[1] Lyu, H., Jiang, S., Zeng, H., Xia, Y., Wang, Q., Zhang, S., & Luo, J. (2023). LLMRec: Personalized Recommendation via Prompting Large Language Models. arXiv:2307.15780. [2] Li, L., Zhang, Y., Liu, D., & Chen, L. (2023). Large Language Models for Generative Recommendation: A Survey and Visionary Discussions. arXiv:2309.01157. [3] Tan, Z., Zeng, Q., Tian, Y., Liu, Z., Yin, B., & Jiang, M. (2024). Democratizing Large Language Models via Personalized Parameter-Efficient Fine-Tuning. arXiv:2402.04401. [4] Wu, L., Zheng, Z., Qiu, Z., Wang, H., Gu, H., Shen, T., Qin, C., Zhu, C., Zhu, H., Liu, Q., Xiong, H., & Chen, E. (2023). A Survey on Large Language Models for Recommendation. arXiv:2305.19860. [5] Netflix Foundation Model Team. (2025). Foundation Model for Personalized Recommendation. Netflix Tech Blog. [6] Spurlock, K. D., Acun, C., Saka, E., & Nasraoui, O. (2024). ChatGPT for Conversational Recommendation: Refining Recommendations by Reprompting with Feedback. arXiv:2401.03605. [7] Dao, H., et al. (2024). Demonstration-augmented Prompt Learning for Conversational Recommendation. SIGIR 2024. [8] Li, L., Zhang, Y., & Chen, L. (2022). Personalized Prompt Learning for Explainable Recommendation. arXiv:2202.07371. [9] Li, Z., Ji, J., Ge, Y., Hua, W., & Zhang, Y. (2024). PAP-REC: Personalized Automatic Prompt for Recommendation Language Models. arXiv:2402.00284 [10] Zou, J., Sun, A., Long, C., & Kanoulas, E. (2024). Knowledge-Enhanced Conversational Recommendation via Transformer-based Sequential Modelling (TSCRKG). ACM Digital Library. [11] Chen, B., Zhang, Z., Langrené, N., & Zhu, S. (2023-2025). Unleashing the Potential of Prompt Engineering for Large Language Models. arXiv:2310.14735. [12] Chen, Q., et al. (2025). Improving Optimal Prompt Learning through Multilayer Integration and Global Attention Mechanisms. Frontiers in Robotics and AI. [13] Ai, J., et al. (2025). An Explainable Recommendation Algorithm Based on Content Summarization and Linear Attention Mechanism. Neurocomputing. [14] Xian, Y., et al. (2019). Reinforcement Knowledge Graph Reasoning for Explainable Recommendation. arXiv:1906.05237. [15] Li, L., Zhang, Y., & Chen, L. (2021). Personalized Transformer for Explainable Recommendation (PETER). ACL 2021, arXiv:2105.11601. [16] Montagna, A. (2023). Graph-Based Explainable Recommendation Systems. CEUR Workshop Proceedings. [17] Rong, Z., et al. (2024). Enhanced Knowledge Graph Recommendation Algorithm Based on Multi-Hop Reasoning and Multi-Level Contrastive Learning. Nature Scientific Reports. [18] Khan, H.U., Naz, A., Alarfaj, F.K., & Almusallam, N. (2025). A Transformer-based Architecture for Collaborative Filtering in Personalized Recommender Systems. Scientific Reports. [19] Chen, F., et al. (2023). KHGCN: Knowledge-Enhanced Hierarchical Graph Capsule Network for Personalized Recommendation. PMC. [20] Zhang, M. (2025). Design and Implementation of Intelligent Library Personalized Recommendation Model Based on Reinforcement Learning. Journal of Combination Mathematics and Computing Informatica.

Copyright

Copyright © 2025 Dr. S Gunasekaran, Ms. Kavitha S, Abhinav A R, Aadith Krishnan K, Reyzon Joseph K R. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.