Abstract
Video-capable multimodal large language models exhibit apparent audio understanding driven by visual cues rather than actual audio processing, necessitating intervention-based frameworks for diagnosing and improving audio-visual alignment.
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.
Community
WVS-Thud: an intervention-driven framework that mitigates the audio-visual Clever Hans effect by teaching video models to verify actual sounds instead of relying on visual shortcuts.
the counterfactual grounding angle in thud is a much-needed move to separate true audio verification from mere visual priors. my worry is ecological validity: beyond shift/mute/swap, do subtle perturbations like brief frequency masking or slight time-stretching reveal more brittle shortcuts that these models still rely on? if they didn't test those, it would be a nice addition, since real-world misalignments are often subtle rather than full breaks. the arxivlens breakdown helped me parse the method details, and i appreciated how it highlights where the grounding signal actually comes from, with a pointer at https://arxivlens.com/PaperView/Details/when-vision-speaks-for-sound-1337-32586847
Get this paper in your agent:
hf papers read 2605.16403 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 4
Rakancorle1/hans-sft-4k
Spaces citing this paper 0
No Space linking this paper