Enregistré dans:
Détails bibliographiques
Auteurs principaux: He, Peize, Wen, Zichen, Wang, Yubo, Wang, Yuxuan, Liu, Xiaoqian, Huang, Jiajie, Lei, Zehui, Gu, Zhuangcheng, Jin, Xiangqi, Yang, Jiabing, Li, Kai, Liu, Zhifei, Li, Weijia, Wang, Cunxiang, He, Conghui, Zhang, Linfeng
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2510.07293
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866912636456665088
author He, Peize
Wen, Zichen
Wang, Yubo
Wang, Yuxuan
Liu, Xiaoqian
Huang, Jiajie
Lei, Zehui
Gu, Zhuangcheng
Jin, Xiangqi
Yang, Jiabing
Li, Kai
Liu, Zhifei
Li, Weijia
Wang, Cunxiang
He, Conghui
Zhang, Linfeng
author_facet He, Peize
Wen, Zichen
Wang, Yubo
Wang, Yuxuan
Liu, Xiaoqian
Huang, Jiajie
Lei, Zehui
Gu, Zhuangcheng
Jin, Xiangqi
Yang, Jiabing
Li, Kai
Liu, Zhifei
Li, Weijia
Wang, Cunxiang
He, Conghui
Zhang, Linfeng
contents Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2510_07293
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs
He, Peize
Wen, Zichen
Wang, Yubo
Wang, Yuxuan
Liu, Xiaoqian
Huang, Jiajie
Lei, Zehui
Gu, Zhuangcheng
Jin, Xiangqi
Yang, Jiabing
Li, Kai
Liu, Zhifei
Li, Weijia
Wang, Cunxiang
He, Conghui
Zhang, Linfeng
Sound
Artificial Intelligence
Computation and Language
Audio and Speech Processing
Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.
title AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs
topic Sound
Artificial Intelligence
Computation and Language
Audio and Speech Processing
url https://arxiv.org/abs/2510.07293