Reproduce “AudioGenie-Reasoner: A Training-Free Multi-Agent Framework for Coarse-to-Fine Audio Deep Reasoning” 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

Yan Rong, Chenxing Li, Dong Yu, Li Liu

The AudioGenie-Reasoner (AGR) framework introduces a training-free multi-agent system that converts audio deep reasoning tasks into iterative text understanding problems, leveraging pre-trained Large Language Models and Audio Large Language Models. This framework achieves state-of-the-art performance on challenging audio reasoning benchmarks by proactively refining an evidence chain.

Key insights

  • Decoupling audio perception from cognitive reasoning enables the system to harness the extensive reasoning capabilities of pre-trained LLMs on text, bypassing the need for specialized audio-reasoning training data.
  • An iterative, multi-agent approach, mimicking human coarse-to-fine cognitive processing, allows for active information seeking and continuous refinement of understanding, crucial for complex and ambiguous audio queries.
  • The 'training-free' nature of orchestrating existing powerful ALLMs and LLMs provides a practical and efficient method for advanced audio reasoning without developing new end-to-end models.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.