1. Introduction

This report provides a detailed analysis of 24 research papers published in the Computer Science - Operating Systems (cs.OS) category on arXiv between March 1, 2026, and March 31, 2026. The goal is to identify emerging research hotspots, understand the author collaboration landscape, and summarize key technological innovations. The analysis reveals a strong convergence between operating systems research and AI/LLM systems, with significant attention to memory management, GPU scheduling, and security in emerging computing paradigms.

2. Research Hotspots and Trends

The papers analyzed cluster around several key areas. We identified seven major research hotspots, which highlight the current priorities and future directions of operating systems research.

3. Author Relationship Graph

The analysis of authorship reveals a diverse research community with no dominant single-author publications. Each paper represents a distinct research team, highlighting the collaborative nature of modern OS research. The following diagrams visualize the collaboration clusters for representative papers.

4. Key Technical Innovations

Across the diverse research topics, several specific, named technologies and frameworks stand out as significant contributions.

AgentRM

An OS-inspired resource manager for LLM agent systems that addresses scheduling failures (blocking, zombie processes) and memory management challenges identified through empirical analysis of 40,000 GitHub issues from six major agent frameworks.

CATS (Conductor-driven Asymmetric Transport Scheme)

A framework that provides TCP with semantic awareness to prioritize critical content, challenging the traditional blind FIFO conveyor belt model of transport protocols.

DetShare

A GPU sharing system that guarantees both semantic determinism (behavioral equivalence) and performance determinism without requiring invasive kernel modifications.

FlexServe

A secure LLM serving system for mobile devices that uses ARM TrustZone with flexible resource isolation, addressing the significant overhead of traditional TrustZone protection for LLM inference.

GateANN

An I/O-efficient SSD-based graph ANNS system supporting filtered vector search on unmodified graph indexes by decoupling graph traversal from vector retrieval.

KRONOS

A software-based, non-intrusive fault injection framework for assessing FreeRTOS dependability against radiation-induced transient and permanent faults.

LMetric

A surprisingly simple multiplication-based metric for LLM request scheduling that achieves both KV cache availability and workload balance, demonstrating that "simple is better" for this complex problem.

Mica

A confidential computing architecture that decouples confidentiality from trust, enabling attestable and trusted workflows composed from independently developed, potentially distrustful TEE components.

NCCLbpf

A verified, high-performance extension framework embedding userspace eBPF runtime into NCCL, bringing kernel-level extensibility safety guarantees to GPU collective communication.

Pichay

A demand paging system for LLM context windows that treats context as L1 cache, introducing proper memory hierarchy (L2, virtual memory, paging) to reduce structural waste identified across 857 production sessions.

Quine

A runtime architecture that realizes LLM agents as native POSIX processes, with explicit mapping: identity is PID, interface is standard streams, state is memory, and environment is inherited.

Structured Gossip DNS

A partition-resilient DNS for MANETs and edge computing that exploits DHT finger tables for passive stabilization, reducing message complexity from O(n) to O(n/log n).

Tock

A research-originated OS that has successfully deployed to secure 10 million computers as root of trust hardware, demonstrating the path from academic research to production-scale security systems.

Wayfinder

An automated OS specialization system using Bayesian optimization over configuration spaces, addressing the challenge of finding optimal configurations among millions of options.

5. Conclusion

The research landscape of operating systems in March 2026 reveals a profound convergence with AI systems. The most striking trend is the application of decades of OS wisdom—resource management, process abstraction, memory hierarchy, scheduling—to LLM agent systems and GPU infrastructure. Papers like AgentRM, Quine, and Pichay demonstrate that the AI community's reinvention of OS primitives at the application layer can benefit from explicit mapping to native OS concepts.

Memory management remains the central challenge, appearing across multiple contexts: tiered memory architectures, mobile app launches, and fundamentally reconceptualized as a hierarchy for LLM context windows. The "Missing Memory Hierarchy" paper is particularly noteworthy for its paradigm-shifting view of LLM context as L1 cache.

Security research continues to push toward zero-trust architectures, with systems designed to minimize the trusted computing base to hardware alone. The journey of Tock from research OS to securing 10 million computers exemplifies the real-world impact possible when security is designed as a first principle.

The innovations presented in this collection demonstrate an adaptive field actively solving the critical systems-level challenges posed by AI workloads, edge computing, and security in an increasingly hostile threat landscape.

6. Full Paper Index