DabDans

HeadRouter

Dynamic Head-Weight Routing for Task-Adaptive Audio Token Pruning in Large Audio Language Models

HeadRouter is a training-free audio token pruning method for large audio language models. It uses task-adaptive head-weight routing to preserve important audio tokens under different semantic and acoustic workloads.

Highlights

• Task-adaptive routing: dynamically adjusts attention-head weights for different audio task types.
• Training-free pruning: applies without additional model training.
• Audio-focused efficiency: targets long-context audio understanding in large audio language models.
• Strong compression performance: retains competitive or improved performance under aggressive token pruning ratios.

Paper

This page accompanies the HeadRouter manuscript and project materials.

• Repository: https://github.com/DabDans/Headrouter_arxiv
• Homepage: https://dabdans.github.io/

Citation

Citation information will be added after release.