ICMOS: Incremental Concept Mining for OS Kernel Configuration via LLMs Agentic Reasoning

ICLR 2026 Conference Submission16299 Authors

Keywords: LLMs, Knowledge Graph, OS kernel configuration, Agentic Reasoning

TL;DR: We propose ICMOS, a framework that combines LLMs with a knowledge graph to enable context-aware concept mining and agentic concept evolution for kernel configuration understanding and optimization.

Abstract: Linux kernel configuration is critical for system performance, security, and adaptability. However, its vast configuration space, comprising over 17,000 configuration options, renders manual tuning both time-consuming and prone to errors. Existing methods largely rely on static heuristics or limited semantic rules, which struggle to capture complex configuration dependencies or adapt across diverse workloads. We introduce ICMOS, a framework that integrates large language models (LLMs) with a heterogeneous knowledge graph of kernel configuration concepts (OSKC-KG). By grounding LLM reasoning in structured semantics, ICMOS supports context-aware mining of configuration concepts and agentic concept evolution in response to new requirements and kernel updates. We evaluate ICMOS on configuration QA tasks and diverse real-world workloads, including databases, web servers, in-memory caches, and system benchmarks. ICMOS consistently outperforms LLM-only baselines, delivering higher accuracy, faster optimization, and robust system performance. Notably, it halves optimization time, reduces tail latency by 58.1\%, and more than doubles configuration success rates. These results demonstrate that ICMOS provides a scalable and reliable framework for grounding LLM reasoning in structured semantics, thereby advancing kernel configuration understanding and optimization.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 16299