Improving Audio Question Answering with Variational Inference

TL;DR

This paper explores the use of variational inference with the IVON optimizer to improve the accuracy and calibration of multimodal large language models in audio question answering tasks, emphasizing more reliable confidence estimates.

Contribution

It demonstrates that variational inference enhances both predictive accuracy and calibration in multimodal models for audio question answering, a novel application of VI in this domain.

Findings

VI improves model calibration and reduces overconfidence.

Application of IVON optimizer enhances fine-tuning performance.

Better confidence estimates support risk-sensitive applications.

Abstract

Variational inference (VI) provides a principled framework for estimating posterior distributions over model parameters, enabling explicit modeling of weight uncertainty during optimization. By capturing this uncertainty, VI improves the reliability of predictions, yielding better calibrated outputs. In this work, we investigate the benefits of VI for challenging multimodal understanding and reasoning by applying the Improved Variational Online Newton (IVON), a recent VI optimizer, to fine-tuning a multimodal large language model on audio question answering tasks. Our results show that VI not only enhances predictive accuracy but also significantly improves calibration, reducing the model's overconfidence. These advances further support risk-sensitive applications such as selective prediction, where reliable confidence estimates are crucial.