AudioMarathon

AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs

Peize He, Zichen Wen, Yubo Wang, Yuxuan Wang, Xiaoqian Liu, Jiajie Huang, et al. · 2025-10-08

Read the paper

Modality

Audio & speech

Data

Hybrid (real + synthetic/simulated)

Task · ground truth

QA/Reason. · Ans.+Eff.

Scope

Broader evaluation with memory demands

What it tests

• CECross-modal evidencePartially evaluated
• LELong-range evidenceDirectly evaluated
• SCState changesNot evaluated
• CFConflicting memoriesNot evaluated
• UQUnanswerable queriesNot evaluated
• MAMemory-grounded actionNot evaluated
• SPSpeaker trackingPartially evaluated
• ACAcoustic-cue retentionDirectly evaluated
• RPRepair persistenceNot evaluated

Survey notes

Long-audio understanding and reasoning are the evaluation targets; memory-efficient model processing is not itself a benchmark of persistent agent memory.