End-to-end speech large language models ((LLMs)) extend the capabilities of text-based models to directly process and generate audio tokens. However, this often leads to a decline in reasoning and generation performance compared to text input, a phenomenon referred to as intelligence degradation. To systematically evaluate this gap, we propose S2SBench, a benchmark designed to quantify performance degradation in Speech LLMs. It includes diagnostic datasets targeting sentence continuation and commonsense reasoning under audio input. We further introduce a pairwise evaluation protocol based on perplexity differences between plausible and implausible samples to measure degradation relative to text input. We apply S2SBench to analyze the training process of Baichuan-Audio, which further demonstrates the benchmark's effectiveness. All datasets and evaluation code are available at https://github.com/undobug/S2SBench.
@article{arxiv.2505.14438,
title = {S2SBench: A Benchmark for Quantifying Intelligence Degradation in Speech-to-Speech Large Language Models},
author = {Yuanbo Fang and Haoze Sun and Jun Liu and Tao Zhang and Zenan Zhou and Weipeng Chen and Xiaofen Xing and Xiangmin Xu},
journal= {arXiv preprint arXiv:2505.14438},
year = {2025}
}