Generate, Analyze, and Refine: Training-Free Sound Source Localization via MLLM Meta-Reasoning

Subin Park, Jung Uk Kim · 2026

Read Generate, Analyze, and Refine: Training-Free Sound Source Localization via MLLM Meta-Reasoning by Subin Park, Jung Uk Kim (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-06824.txt.

Description

Park and Kim localize sound sources without training, using multimodal models that reason about their own work. Most sound-source finding leans on contrastive matching, lacking explicit checks in complex noisy scenes. Their three-stage pipeline first generates boxes and audio labels, then scores audio-visual consistency through open-set role tagging and anchor voting, and finally gates tweaks on the fly to avoid extra changes. Trials on single-source and multi-source benchmarks show competitive results, with code shared.

Catalog record

CREATORSubin Park, Jung Uk Kim
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45881
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.06824

Browse the full library · About the Tower of Babel