TLDR: A new benchmark, MCR-BENCH, reveals that Large Audio-Language Models (LALMs) consistently prioritize textual input over audio evidence when presented with conflicting information, leading to significant performance degradation. Despite internal awareness of inconsistencies, LALMs remain overconfident. While prompting offers limited help, supervised finetuning shows more promise in mitigating this text bias, but a complete solution requires deeper architectural changes.
Large Audio-Language Models, or LALMs, are advanced AI systems designed to understand and process both audio and text inputs. These models have shown impressive capabilities in various tasks like answering questions about audio, detecting sound events, and recognizing speech. However, a new research paper highlights a significant challenge: how these models perform when the audio and text information they receive contradict each other.
The paper, titled “When Audio and Text Disagree: Benchmarking Text Bias in Large Audio-Language Models under Cross-Modal Inconsistencies,” introduces a new benchmark called MCR-BENCH. This benchmark is the first of its kind specifically designed to evaluate how LALMs prioritize information when presented with conflicting audio-text pairs. The researchers, Cheng Wang, Gelei Deng, Xianglin Yang, Han Qiu, and Tianwei Zhang, found a concerning trend: LALMs exhibit a strong bias towards textual input, often ignoring audio evidence when inconsistencies arise. This bias leads to a significant drop in performance for audio-focused tasks and raises serious questions about the reliability of these models in real-world applications.
To understand this problem, the MCR-BENCH benchmark includes 3,000 specially created samples across three audio understanding tasks: Audio Question Answering (AQA), Speech Emotion Recognition (SER), and Vocal Sound Classification (VSCn). For each audio sample, three types of textual descriptions are generated:
Faithful Text
These are accurate descriptions that correctly represent the audio content.
Adversarial Text
These are deliberately misleading descriptions that directly contradict the audio content.
Also Read:
- LLaSO: An Open Standard for Large Speech-Language Models
- The Hidden Challenge of Noisy Data in Speech Separation
Irrelevant Text
These descriptions are semantically unrelated to the audio content.
The evaluation of several state-of-the-art LALMs on MCR-BENCH revealed a consistent pattern of text bias. When models were given adversarial textual descriptions, their performance dropped dramatically. For example, in the Audio Question Answering task, some models saw their accuracy plummet from over 87% to less than 2%, representing a performance deterioration of more than 98%. In the Speech Emotion Recognition task, some models even showed 0% accuracy when faced with contradictory text. This indicates that LALMs overwhelmingly favor text when trying to resolve conflicts between modalities.
While this text bias was universal, some models demonstrated slightly better resilience. Audio-Flamingo2, for instance, showed stronger robustness, maintaining higher adversarial accuracy compared to its competitors. However, even the most robust models still showed considerable vulnerability to text bias. The study also found that the severity of text bias varied across different tasks, with emotion recognition tasks being particularly susceptible to misleading text.
Further investigation into the causes of this text bias revealed interesting insights. LALMs maintain remarkably high confidence scores even when making incorrect predictions due to adversarial text. This suggests that the models lack effective mechanisms to detect and appropriately respond to cross-modal inconsistencies. Despite this overconfidence, internal analysis showed that LALMs do internally recognize inconsistencies between audio and text. However, this internal awareness doesn’t translate into appropriate decision-making or output behavior, highlighting a disconnect between what the model “knows” and what it “says.”
The researchers explored mitigation strategies to address this text bias. Simple prompting techniques, such as instructing the model to prioritize audio or be aware of potential biases, offered only limited improvements. A more promising approach was Supervised Finetuning (SFT) on datasets specifically designed with conflicting audio-text pairs. This method significantly improved adversarial accuracy and reduced the text’s influence, though it sometimes came with a slight trade-off in performance on faithful text conditions. While SFT showed better results, it did not completely eliminate the text bias, indicating that more fundamental architectural or training modifications might be necessary.
In conclusion, the MCR-BENCH benchmark underscores a critical challenge in the development of robust LALMs. The strong bias towards textual input when audio and text conflict raises significant reliability concerns for real-world applications. The findings highlight the urgent need for new training methods and more sophisticated ways for these models to integrate and balance information from different modalities. This research provides a valuable foundation for developing more dependable large audio-language models in the future. You can read the full research paper here: When Audio and Text Disagree: Benchmarking Text Bias in Large Audio-Language Models under Cross-Modal Inconsistencies.
