When Vision Speaks for Sound

TL;DR

This paper reveals that current video-capable multimodal models often rely on visual cues to infer audio information rather than verifying audio streams, highlighting a vision-driven bias.

Contribution

The authors introduce Thud, a probing framework with counterfactual audio edits, and propose a two-stage alignment method to improve audio verification in multimodal models.

Findings

Models exploit visual-acoustic correlations without true audio verification.

The proposed method improves performance on intervention dimensions by 28 percentage points.

Slight improvements observed on general video and audio-visual QA benchmarks.

Abstract

Despite rapid progress in video-capable MLLMs, we find that their apparent audio understanding in videos is often vision-driven: models rely on visual cues to infer or hallucinate acoustic information, rather than verifying the audio stream. This issue appears across both state-of-the-art open-source omni models and leading closed-source models from providers such as Google and OpenAI. We characterize this failure mode as an audio-visual Clever Hans effect, in which models appear (falsely) audio-grounded, but actually exploit visual-acoustic correlations without verifying whether the audio and visual streams are truly aligned. To systematically study this behavior, we introduce Thud, an intervention-driven probing framework based on three counterfactual audio edits: Shift, which tests temporal synchronization; Mute, which tests sound existence; and Swap, which tests audio-visual consistency. Beyond diagnosis, we further study a two-stage alignment recipe: intervention-derived preference pairs teach audio verification, while event-level general video preferences regularize the model against over-specialization. Our best 10K-sample recipe improves average performance across the three intervention dimensions by 28 percentage points, while slightly improving performance on general video and audio-visual QA benchmarks.