Research map / Large language models / Agents, code and math
OSVBench: Benchmarking LLMs on Specification Generation Tasks for Operating System Verification
Shangyu Li, Juyong Jiang, Tiancheng Zhao, Jiasi Shen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add few-shot syscall examples (0-shot to 1/3/5-shot) (ablation) | five LLMs on OSVBench · OSVBench | Pass@1: 0% at 0-shot, large gains with more shots | Improved |
| increase number of demonstrations from 1 to 5 (ablation) | DeepSeek-Chat, o3-mini and others · OSVBench | Pass@1 | Improved |
| inject bugs into the syscall code in the prompt (ablation) | 12 LLMs · OSVBench | Pass@1: 53 of 60 (88.3%) bug experiments lower than bug-free | Hurt |
| increase number of injected bugs per task (ablation) | 12 LLMs · OSVBench | Pass@1 | Hurt |
| use reasoning model (o1) instead of GPT-4o | OpenAI models · OSVBench | Pass@1 total: 38.78 -> 23.67 | Hurt |
| use reasoning model DeepSeek-R1 instead of DeepSeek-Chat | DeepSeek · OSVBench | Pass@1 total: 46.53 -> 40.82 | Hurt |
| use larger model (70B/72B vs 7B/8B) | Llama-3.1 and Qwen2.5 · OSVBench | Pass@1 total: 10.61 -> 22.45 (Llama); 4.90 -> 32.24 (Qwen) | Improved |
| increase k in Pass@k from 1 to 3 to 5 (ablation) | GPT-4o, DeepSeek-Chat, Llama-3.1-70B-Instruct · OSVBench | Pass@k | Improved |
| add self-repair with verifier error messages (ablation) | GPT-4o and DeepSeek-Chat · OSVBench | Repair success rate: 14.67 (round 1), 1.56 (round 2) for GPT-4o | Improved |
| add a second self-repair round (ablation) | GPT-4o and DeepSeek-Chat · OSVBench | Repair success rate: 14.67 -> 1.56 (GPT-4o); 11.45 -> 2.59 (DeepSeek-Chat) | No clear effect |
Closest papers by meaning
- CoRe: Benchmarking LLMs’ Code Reasoning Capabilities through Static Analysis Tasks (NeurIPS 2025 · similarity 0.87)
- kGym: A Platform and Dataset to Benchmark Large Language Models on Linux Kernel Crash Resolution (NeurIPS 2024 · similarity 0.86)
- AlphaVerus: Bootstrapping Formally Verified Code Generation through Self-Improving Translation and Treefinement (ICML 2025 · similarity 0.86)
- Lemur: Integrating Large Language Models in Automated Program Verification (ICLR 2024 · similarity 0.86)
- HierSVA: A Synthesis Pipeline, Dataset, and Benchmark for LLM-Driven Hierarchical Hardware Formal Verification (NeurIPS 2026 · similarity 0.86)
- How efficient is LLM-generated code? A rigorous & high-standard benchmark (ICLR 2025 · similarity 0.86)
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.
