Research map / Large language models / Agents, code and math
BEST: Benchmarking Efficiency in Space and Time for LLM-Generated Code
Aocheng Shen, Boyu Zhang, Jiaze Li, Ruixuan Ma, Qiankun Zhang, Jing Wang, Bin Yuan, Shenghao Liu, Xianjun Deng
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 |
|---|---|---|---|
| measure dual efficiency (time and space) instead of correctness only | 50 LLMs on BEST · BEST | dual@10 vs pass@10: GPT-5.2: pass@10 0.879 -> dual@10 0.811 | Hurt |
| use reasoning (thinking) mode in the LLM (ablation) | DeepSeek-V3.2 · BEST (tau=sigma=1.2) | dual@10: 0.630 -> 0.678 | Improved |
| use reasoning (thinking) mode in the LLM (ablation) | DeepSeek-V3.2 · BEST (tau=sigma=1.2) | dual@1: 0.510 -> 0.508 | No clear effect |
| measure space-only efficiency instead of time-only efficiency | LLMs generally · BEST | dual@10: Llama-3.1-8B time-only 0.244 vs space-only 0.248 (the only exception) | Hurt |
| sample 10 codes instead of 1 (temperature 0.8) (ablation) | all LLMs average · BEST | pass@k: 0.351 -> 0.495 | Improved |
| sample 10 codes instead of 1 (temperature 0.8) (ablation) | all LLMs average · BEST | dual@k: 0.250 -> 0.349 | Improved |
| newer model release within DeepSeek family | DeepSeek models · BEST (tau=sigma=1.2) | dual@10: DeepSeek-Coder-33B-Instruct 0.246 -> DeepSeek-V3.2-Thinking 0.678 | Improved |
| larger model size within the Llama family (ablation) | Llama-3.1 · BEST (tau=sigma=1.2) | dual@10: 0.194 -> 0.410 | Improved |
| use instruction-tuned instead of base variant (ablation) | DeepSeek-Coder-33B · BEST (tau=sigma=1.2) | dual@10: 0.216 -> 0.246 | Improved |
| better correctness of the model | CodeGemma-7B vs Phind-CodeLlama-34B-V1 · BEST | dual@10 | Mixed |
Closest papers by meaning
- How efficient is LLM-generated code? A rigorous & high-standard benchmark (ICLR 2025 · similarity 0.95)
- EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code (NeurIPS 2025 · similarity 0.92)
- Mercury: A Code Efficiency Benchmark for Code Large Language Models (NeurIPS 2024 · similarity 0.91)
- LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code (ICLR 2025 · similarity 0.90)
- StackEval: Benchmarking LLMs in Coding Assistance (NeurIPS 2024 · similarity 0.90)
- Let's Revise Step-by-Step: A Unified Local Search Framework for Code Generation with LLMs (NeurIPS 2025 · similarity 0.89)
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.
