Reproduce “SAVU-BENCH: A Real-World Benchmark for Spatial Audio-Visual Understanding” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yu Chen, Ruihang Liu, Yangguang Xu, Xinyue Jiang, Mohammed Bennamoun, Farid Boussaid, Xinyuan Qian, Qiuhong Ke

The authors introduce SAVU-Bench to evaluate spatial audio-visual perception, alignment, and reasoning on real-world recordings. It also links reasoning questions to diagnostic tests of their grounding and alignment prerequisites. With predicted spatial cues, SAVU-EA raises Gemini-3.6-Flash’s audio grounding from 52.5% to 83.2% and spatial reasoning from 44.8% to 59.4%. Correct oracle facts improve reasoning, but do not eliminate errors.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.