Investigating Faithfulness in Large Audio Language Models

A new study questions whether the reasoning steps generated by Large Audio Language Models are truly grounded in the audio they process. Researchers have proposed a systematic framework to evaluate the faithfulness of Chain-of-Thought explanations in these multimodal systems, identifying a potential disconnect between fluent reasoning and actual auditory input. Large Audio Language Models, or LALMs, combine audio encoders with pretrained Large Language Models to handle tasks that require both listening and reasoning [1]. While they can produce Chain-of-Thought explanations — step-by-step rationales for their conclusions — the reliability of these chains has not been systematically verified [2]. A research team led by Cem Subakan has introduced a framework to measure this faithfulness, defining three core criteria: the reasoning must be hallucination-free, holistic, and demonstrate attentive listening [2]. In artificial intelligence, a hallucination is a response that contains false or misleading information presented as fact [3]. Such errors pose significant challenges for deploying large language models in high-stakes fields like medical diagnostics and supply chain logistics [3]. The new benchmark applies this concern to the audio domain by introducing interventions on both the audio input and the reasoning chain itself to test whether the model's explanations hold up under scrutiny [2]. The framework was tested on two models, Audio Flamingo 3 and Qwen2.5-Omni [2]. Results pointed to a multimodal disconnect: the generated reasoning often aligned with the model's final prediction but was not always strongly anchored in the actual audio content [2]. The models proved vulnerable to hallucinations and could be swayed by adversarial perturbations, raising questions about their trustworthiness in real-world applications [2]. The work arrives as the broader AI field grapples with the reliability of large language models, which are statistical systems trained on vast text corpora [11]. Unlike symbolic AI, which generally avoids hallucination, these neural network-based models can embed plausible-sounding falsehoods in their outputs [3]. The study extends this line of inquiry into the audio modality, where a model might, for example, describe a sound that is not present in the recording. The paper was submitted to the arXiv preprint repository on September 26, 2025, and revised through four versions, with the latest posted on June 17, 2026 [1]. arXiv, which began in 1991, now hosts over two million e-prints and receives roughly 24,000 new submissions each month, serving as a primary venue for rapid dissemination in computer science and physics [9]. The researchers have also released a benchmarking interface and evaluation results on a dedicated project page [2].