Research map / Large language models / Agents, code and math
AutoOS: Make Your OS More Powerful by Exploiting Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
AutoOS uses GPT-3.5-Turbo in a state-machine observe-prune-propose-act-correct loop over a dynamic kconfig tree to customize Linux kernel configurations for better UnixBench performance while ensuring boot-up.
Problem
Automatically customizing and optimizing Linux kernel configurations (over 15,000 options, 1-2 hour evaluations, error-prone options causing boot failure) for AIoT and general-purpose scenarios.
Approach
Formulates kernel configuration as optimization on a dynamic tree and uses an LLM as a state-machine controller with randomized pruning and proposing, self-explanation for consistency, and LLM-based boot-failure correction, interacting with the kernel via kconfiglib-based atomic APIs.
Key results
- On PolyOS (AIoT, HiFive board), AutoOS reaches a UnixBench total score of 335 vs 309 for the vendor/expert default (+8.4%), and +25.6% on Fedora (260 vs 207).
- On Ubuntu on a PC, AutoOS improves UnixBench total score by 9.08% (4238 vs 3885) over default, while LLM-Vanilla gives only +0.3%.
- Self-explanation with 2 iterations raises consistency from 28.6% to 93.38%, and removes boot-up failure options on PolyOS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AutoOS instead of the vendor default configuration | PolyOS · SiFive Unmatched board, UnixBench | total score: 309 -> 335 (+8.4%) | Helped |
| Directly prompt an LLM to recommend configuration options instead of using the default | PolyOS · SiFive Unmatched board, UnixBench | total score: 309 -> 283 (-8.5%) | Hurt |
| Use AutoOS instead of the vendor default configuration | Fedora · SiFive Unmatched board, UnixBench | total score: 207 -> 260 (+25.6%) | Helped |
| Directly prompt an LLM to recommend configuration options instead of using the default | Fedora · SiFive Unmatched board, UnixBench | total score: 207 -> 194 (-6.3%) | Hurt |
| Use AutoOS instead of the vendor default configuration | Ubuntu · PC, UnixBench | total score: 3885 -> 4238 (+9.0%) | Helped |
| Directly prompt an LLM to recommend configuration options instead of using the default | Ubuntu · PC, UnixBench | total score: 3885 -> 3898(+0.3%) | No clear effect |
| Add pruning and self-explanation to plain AutoOS (ablation) | AutoOS · PolyOS on AIoT device, UnixBench | total score: 137 -> 335 | Helped |
| Add self-explanation to AutoOS with pruning (ablation) | AutoOS · PolyOS on AIoT device, UnixBench | total score: 298 -> 335 | Helped |
| Add pruning to AutoOS with self-explanation (ablation) | AutoOS · PolyOS on AIoT device, UnixBench | total score: 332 -> 335 | Helped |
| Use 1 instead of 0 self-explanation iterations (ablation) | AutoOS · PolyOS configuration traversal | option impact consistency: 28.60% -> 83.08% | Helped |
| Use 2 instead of 1 self-explanation iterations (ablation) | AutoOS · PolyOS configuration traversal | option impact consistency: 83.08% -> 93.38% | Helped |
| Use 8 instead of 2 self-explanation iterations (ablation) | AutoOS · PolyOS configuration traversal | option impact consistency: 93.38% -> 94.85% | No clear effect |
Surprising findings
- Directly prompting ChatGPT/LLM-Vanilla for kernel options performs worse than the default configuration on PolyOS and Fedora (-8.5% and -6.3% total score).
- Pruning plus self-explanation reduces modified options from 230 to 20 and 17 respectively on PolyOS, while improving the total score from 137 to 335 (with self-explanation alone reaching 332).
Limitations
- Currently only applicable to Linux-based OSes or those with kconfig-like configuration commands; non-Linux systems need APIs built first.
- Gains on integer and floating-point sub-scores are limited because they depend on CPU hardware rather than the OS.
- Evaluation relies on a single LLM (GPT-3.5-Turbo) and a limited set of OS distributions and hardware platforms.
Experimental setup
Benchmarks: UnixBench LMbench
Metrics: UnixBench total score LMbench sub-metric performance consistency (%) number of modified options boot-up failure options search time
Baselines: vendor default configuration LLM-Vanilla Genkernel
Builds on: kconfiglib self-explanation
Closest papers by meaning
- kGym: A Platform and Dataset to Benchmark Large Language Models on Linux Kernel Crash Resolution (NeurIPS 2024 · similarity 0.80)
- MigGPT: Harnessing Large Language Models for Automated Migration of Out-of-Tree Linux Kernel Patches Across Versions (NeurIPS 2025 · similarity 0.79)
- OSCAR: Operating System Control via State-Aware Reasoning and Re-Planning (ICLR 2025 · similarity 0.79)
- Outrunning LLM Cutoffs: A Live Kernel Crash Resolution Benchmark for All (ICML 2026 · similarity 0.78)
- macOSWorld: A Multilingual Interactive Benchmark for GUI Agents (NeurIPS 2025 · similarity 0.76)
- EGG: An Expert-Guided Agent Framework for Kernel Generation (ICML 2026 · similarity 0.76)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.