Reproduce “When Sounds Turn Heads: Recovering Event Timing That Audio LLMs Fail to Say” 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
Jongbhin Woo, Jaewook Jeong, Joon Son Chung
The authors use selected attention heads to revise audio language models’ timestamps without changing model weights. For a skidding query, Qwen2.5-Omni’s answer moves from 10.0–10.5 seconds to 14.0–18.0 seconds, within the annotated event. On RealDESED, Qwen2-Audio’s recording mIoU rises from 0.066 to 0.480, versus 0.278 with random heads; revision applies only when the model detects the sound.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.