Evaluation of LLMs Should Not Ignore Non-Determinism

GitHub | 📖 arXiv

Official sampling results for The Good, The Bad, and The Greedy: Evaluation of LLMs Should Not Ignore Non-Determinism

Dataset Details

Authors: Yifan Song, Guoyin Wang, Sujian Li, Bill Yuchen Lin.

Current evaluations of large language models (LLMs) often overlook non-determinism, typically focusing on a single output per example. This limits our understanding of LLM performance variability in real-world applications. Our study addresses this issue by exploring key questions about the performance differences between greedy decoding and sampling, identifying benchmarks’ consistency regarding non-determinism, and examining unique model behaviors.

Here are our findings:

• A notable performance gap is observed between greedy decoding and sampling generation.
• Greedy decoding outperforms sampling on most evaluated benchmarks, except for AlpacaEval.
• Math reasoning and code generation were most impacted by sampling variance.
• The above findings remain consistent across different sizes and families of LLMs.
• Alignment methods, e.g., DPO, can significantly reduce the sampling variance for most benchmarks.
• High temperature will significantly harm the reasoning and code generation capabilities of LLMs, while higher repetition penalty leads to improved performance on AlpacaEval.
• In the best-of-N sampling setting, 7B-level LMs have the potential to outperform GPT-4-Turbo.

We evaluate non-determinism generation of LLMs on seven benchmarks: AlpacaEval 2, Arena-Hard, WildBench v2, MixEval, MMLU-Redux, GSM8K, and HumanEval.

From the results, we observe a consistent performance gap between greedy decoding and the sampling method. Greedy decoding generally proves more effective for most tasks, except for AlpacaEval.

📖 Citation

If you find this repo helpful, please cite out paper:

@article{song2024good,
    author={Yifan Song and Guoyin Wang and Sujian Li and Bill Yuchen Lin},
    title={The Good, The Bad, and The Greedy: Evaluation of LLMs Should Not Ignore Non-Determinism},
    year={2024},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

arxiv id: 2407.10457