Researchers increase Qwen2-Audio accuracy from 57.1% to 71.9%
The grounding profile of one model does not move to a different model.
Claimed, not confirmed
Large audio-language models frequently tell that a sound is in the audio when it is not. This is object hallucination. The researchers made a black-box causal test. It finds the evidence region with CLAP similarity and removes it. Then it compares the result with a random-window control on 3 models. Contrastive decoding with 2 contrasts increases the accuracy of Qwen2-Audio from 57.1% to 71.9%.
Sources
Posted
