arXivPaperNeeds Review
Encoder-Side Neuron Identification and Amplification for Acoustic Perception in Large Audio-Language Models
IAAN boosts selected audio-encoder neurons at inference, improving fine-grained speech perception across three models without retraining or labels.
arXiv
Source Summary
IAAN scores audio-encoder neurons by contrasting responses to a real waveform and a noise reference, then amplifies a small selected set during inference. It requires **no training or labels**.
Practical Implication
Across ten non-semantic speech attributes, average accuracy rose by **25.7 points on Audio-Flamingo-3**, **21.4 on Qwen2.5-Omni**, and **9.7 on Kimi-Audio**. Builders with encoder access could test targeted activation changes before committing to fine-tuning.
Agent-Ready Context
IAAN scores audio-encoder neurons by contrasting responses to a real waveform and a noise reference, then amplifies a small selected set during inference. It requires **no training or labels**. Across ten non-semantic speech attributes, average accuracy rose by **25.7 points on Audio-Flamingo-3**, **21.4 on Qwen2.5-Omni**, and **9.7 on Kimi-Audio**. Builders with encoder access could test targeted activation changes before committing to fine-tuning. The gains depended on intervening inside the encoder and choosing the right neurons. Post-encoder or language-model interventions offered little benefit or reduced accuracy, and the material does not establish results beyond the tested models and attributes.
Context Map
modelaudio#generative-mediaUncertainty
The gains depended on intervening inside the encoder and choosing the right neurons. Post-encoder or language-model interventions offered little benefit or reduced accuracy, and the material does not establish results beyond the tested models and attributes.