Research map / Large language models / Agents, code and math
CIRBench: Evaluating Large Language Models as LLVM IR Optimizers
Zi Yang, Haifeng Ding, Fei Liu, Yingying Cheng, Han Cheng, Zhilei Chai, Haojie Zhou
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 verifier error message and location as a hint (ablation) | GPT-5 · CIRBench Repair | Equiv@1 (%): 74.0 -> 85.5 | Improved |
| add verifier error message and location as a hint (ablation) | Claude Sonnet 4.5 · CIRBench Repair | Equiv@1 (%): 66.0 -> 76.5 | Improved |
| add verifier error message and location as a hint (ablation) | Gemini 2.5 Flash · CIRBench Repair | Equiv@1 (%): 44.0 -> 74.5 | Improved |
| add verifier error message and location as a hint (ablation) | Deepseek-V3.2-Exp · CIRBench Repair | Equiv@1 (%): 39.0 -> 67.0 | Improved |
| add verifier hints on hard and multi-error repairs (ablation) | strong models · CIRBench Repair (Medium/Hard, multi-error) | Equiv@1: 10–30% | Improved |
| sample 5 candidates instead of 1 | GPT-5 · CIRBench Transform | Equiv (%): 73.3 -> 92.7 | Improved |
| sample 5 candidates instead of 1 | Deepseek-V3.2-Exp · CIRBench Transform | Equiv (%): 15.3 -> 40.7 | Improved |
| sample 5 candidates instead of 1 | GPT-5 · CIRBench Refactor-Normal | Equiv (%): 22.5 -> 29.5 | Improved |
| sample 5 candidates instead of 1 | GPT-5 · CIRBench Refactor-Normal | EM (%): 0.0 -> 0.0 | No clear effect |
| run LLM output through -O3 (copilot mode) (ablation) | GPT-5 · CIRBench Transform (perf) | median dynamic speedup vs -O3: 0.85 -> 0.99 | Improved |
| run LLM output through -O3 (copilot mode) (ablation) | Grok 4 Fast · CIRBench Transform (perf) | median dynamic speedup vs -O3: 0.66 -> 0.98 | Improved |
| run LLM output through -O3 (copilot mode) (ablation) | GPT-5 · CIRBench Transform (perf) | max dynamic speedup vs -O3: 4.89 -> 3.23 | Hurt |
Closest papers by meaning
- Can Large Language Models Understand Intermediate Representations in Compilers? (ICML 2025 · similarity 0.91)
- How efficient is LLM-generated code? A rigorous & high-standard benchmark (ICLR 2025 · similarity 0.89)
- EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code (NeurIPS 2025 · similarity 0.88)
- CoRe: Benchmarking LLMs’ Code Reasoning Capabilities through Static Analysis Tasks (NeurIPS 2025 · similarity 0.88)
- REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving (NeurIPS 2025 · similarity 0.88)
- LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code (ICLR 2025 · similarity 0.87)
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.
