Reproduce “STEMMA: Song-to-Stem Multi-Audio Reasoning for Large Audio Language Models” 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
Hoyeol Sohn, Wonil Kim, Keunhyoung Kim, Sangeun Kum, Taehyoung Kim, Dongjoo Moon, Theerasak Charoenchob, Teeratep Weerapang and 2 more
STEMMA builds music-comparison questions from cataloged song, section, and stem provenance, retrieving related excerpts and negatives instead of relying on random sampling. Music Flamingo fine-tuned on STEMMA-Instruct answers four-candidate structural questions with 81.89% accuracy, versus 50.39% for the strongest caption cascade. This result is specific to structural questions; attribute questions remain difficult, with a best score of 35.26%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.