When Vision Speaks for Sound

Hugging Face Daily Papers Papers

Summary

This paper identifies that video-capable multimodal LLMs often appear to understand audio but actually rely on visual cues, a failure mode termed the audio-visual Clever Hans effect. It introduces Thud, an intervention-driven probing framework to diagnose this issue, and proposes an alignment recipe that improves audio-visual consistency by 28 percentage points.

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.
Original Article
View Cached Full Text

Cached at: 05/20/26, 06:36 AM

Paper page - When Vision Speaks for Sound

Source: https://huggingface.co/papers/2605.16403

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 invideo-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 anaudio-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, anintervention-driven probing frameworkbased on threecounterfactual audio edits: Shift, which teststemporal synchronization; Mute, which testssound existence; and Swap, which testsaudio-visual consistency. Beyond diagnosis, we further study a two-stagealignment recipe: intervention-derivedpreference pairsteach audio verification, whileevent-level general video preferencesregularize 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.

View arXiv pageView PDFProject pageGitHub1Add to collection

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 paper0

No model linking this paper

Cite arxiv.org/abs/2605.16403 in a model README.md to link it from this page.

Datasets citing this paper4

#### Rakancorle1/hans-10k Viewer• Updatedabout 15 hours ago • 20.8k • 15 #### Rakancorle1/hans-sft-4k Viewer• Updated1 day ago • 3.83k • 7 #### Rakancorle1/vggsync-3k Viewer• Updated1 day ago • 3k #### Rakancorle1/thud-eval Viewer• Updated1 day ago • 710

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.16403 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Do Vision-Language Models Truly Perform Vision Reasoning? A Rigorous Study of the Modality Gap

arXiv cs.CL

This paper introduces CrossMath, a controlled multimodal reasoning benchmark that reveals a critical limitation in current vision-language models: they perform reasoning primarily in textual space rather than genuine vision-grounded reasoning, with visual input often degrading performance compared to text-only baselines. The authors propose fine-tuning approaches to mitigate this modality gap and improve multimodal reasoning capabilities.