Improving Audio Question Answering with Variational Inference
Haolin Chen
TL;DR
This work tackles the reliability of multimodal audio question answering by addressing weight uncertainty through variational inference. It applies the Improved Variational Online Newton (IVON) optimizer to fine-tune a state-of-the-art multimodal LLM (Qwen2.5-Omni 3B) with LoRA on the DCASE 2025 AQA dataset, evaluating both accuracy and calibration. Results show that IVON yields better-calibrated predictions and substantially improves selective prediction, with MC-based inference (IVON MC-8) offering the strongest reliability benefits. The study demonstrates practical gains for risk-sensitive applications and discusses the trade-offs between MC sampling, posterior variance, and computational overhead, pointing to future work on open-ended generation tasks.
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.