Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Reliable Large Audio Language Model

Ziyang Ma, Xiquan Li, Yakun Song, Wenxi Chen, Chenpeng Du, Jian Wu, Yuanzhe Chen, Zhuo Chen, Yuping Wang, Yuxuan Wang, Xie Chen

TL;DR

This work investigates reliability in large audio language models (LALMs) across speech, music, and general sound. It systematically compares training-free methods (IDK prompting, MCoT prompting, and a Task Agent) with training-based approaches (IDK dataset construction and post-training) and introduces the Reliability Gain Index (RGI) to quantify reliability gains. The study demonstrates that reliability awareness is a transferable meta-ability across modalities, with training-based methods offering a better accuracy-truthfulness trade-off and cross-modal transfer of reliability knowledge, as evidenced by consistent positive RGI across sound, music, and speech. The findings advance reliable LALMs and provide practical metrics and methods for deploying trustworthy audio-language reasoning systems, with MMAU as a benchmark for evaluation.

Abstract

Recent advancements in large audio language models (LALMs) have demonstrated impressive results and promising prospects in universal understanding and reasoning across speech, music, and general sound. However, these models still lack the ability to recognize their knowledge boundaries and refuse to answer questions they don't know proactively. While there have been successful attempts to enhance the reliability of LLMs, reliable LALMs remain largely unexplored. In this paper, we systematically investigate various approaches towards reliable LALMs, including training-free methods such as multi-modal chain-of-thought (MCoT), and training-based methods such as supervised fine-tuning (SFT). Besides, we identify the limitations of previous evaluation metrics and propose a new metric, the Reliability Gain Index (RGI), to assess the effectiveness of different reliable methods. Our findings suggest that both training-free and training-based methods enhance the reliability of LALMs to different extents. Moreover, we find that awareness of reliability is a "meta ability", which can be transferred across different audio modalities, although significant structural and content differences exist among sound, music, and speech.

Towards Reliable Large Audio Language Model

TL;DR

This work investigates reliability in large audio language models (LALMs) across speech, music, and general sound. It systematically compares training-free methods (IDK prompting, MCoT prompting, and a Task Agent) with training-based approaches (IDK dataset construction and post-training) and introduces the Reliability Gain Index (RGI) to quantify reliability gains. The study demonstrates that reliability awareness is a transferable meta-ability across modalities, with training-based methods offering a better accuracy-truthfulness trade-off and cross-modal transfer of reliability knowledge, as evidenced by consistent positive RGI across sound, music, and speech. The findings advance reliable LALMs and provide practical metrics and methods for deploying trustworthy audio-language reasoning systems, with MMAU as a benchmark for evaluation.

Abstract

Recent advancements in large audio language models (LALMs) have demonstrated impressive results and promising prospects in universal understanding and reasoning across speech, music, and general sound. However, these models still lack the ability to recognize their knowledge boundaries and refuse to answer questions they don't know proactively. While there have been successful attempts to enhance the reliability of LLMs, reliable LALMs remain largely unexplored. In this paper, we systematically investigate various approaches towards reliable LALMs, including training-free methods such as multi-modal chain-of-thought (MCoT), and training-based methods such as supervised fine-tuning (SFT). Besides, we identify the limitations of previous evaluation metrics and propose a new metric, the Reliability Gain Index (RGI), to assess the effectiveness of different reliable methods. Our findings suggest that both training-free and training-based methods enhance the reliability of LALMs to different extents. Moreover, we find that awareness of reliability is a "meta ability", which can be transferred across different audio modalities, although significant structural and content differences exist among sound, music, and speech.

Paper Structure

This paper contains 34 sections, 18 equations, 5 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Large Audio Language Model
  4. Evaluation of Reliable Generation
  5. Reliable LALM
  6. Training-free Method
  7. IDK Prompting.
  8. MCoT Prompting.
  9. Task Agent.
  10. Training-based Method
  11. Construction of the IDK Dataset.
  12. Post-training with the IDK Dataset.
  13. Evaluation
  14. Reliability Evaluation
  15. Reliability Gain
  16. ...and 19 more sections

Figures (5)

  • Figure 1: Illustration of Unreliable vs. Reliable LALMs.
  • Figure 2: Illustration of Reliability Awareness Transferability in LALM.
  • Figure 3: Heatmap for cross-modal SFT results. The figure shows the RGI performance of reliable models trained on one modality and tested on another.
  • Figure 4: Rejection Rate(%), Reliability(%), and Reliability Gain Index (RGI) performance for different LoRA alpha weights trained on speech modality.
  • Figure 5: IDK percentage for constructing IDK dataset with different $K$@$5$ threshold.