Do Audio-Visual Large Language Models Really See and Hear?

Ramaneswaran Selvakumar; Kaousheik Jayakumar; S Sakshi; Sreyan Ghosh; Ruohan Gao; Dinesh Manocha

Source-linked AI summary

Do Audio-Visual Large Language Models Really See and Hear?

Ramaneswaran Selvakumar, Kaousheik Jayakumar, S Sakshi, Sreyan Ghosh, Ruohan Gao, Dinesh Manocha

arXiv:2604.02605v1Apr 3, 2026cs.AIcs.SD

TL;DR

AVLLMs can process audio and vision, but their internal integration and reliability under conflicting modalities are not well understood. The paper conducts the first mechanistic interpretability study of AVLLMs across layers, using attention analysis, representation probing, attention knockouts, and comparisons with base vision-language models. It finds that latent audio understanding is present but is suppressed during generation as deeper layers favor vision, with this bias potentially linked to vision-heavy training.

Problem
AVLLM audio-visual integration remains mechanistically opaque, especially when conflicting modalities raise reliability concerns.
Method
The paper analyzes attention and audio representations across layers, applies attention knockouts, and compares AVLLM token distributions with base vision-language models.
Results
Audio understanding deteriorates by up to 56% under conflicting audio and visual cues, despite latent audio semantics; deeper layers favor vision and suppress audio.
Takeaways & Limitations
AVLLMs systematically prefer visual cues even for audio understanding, supporting counterfactual evaluation and training approaches that address modality imbalance.

Abstract

from arXiv · show

Audio-Visual Large Language Models (AVLLMs) are emerging as unified interfaces to multimodal perception. We present the first mechanistic interpretability study of AVLLMs, analyzing how audio and visual features evolve and fuse through different layers of an AVLLM to produce the final text outputs. We find that although AVLLMs encode rich audio semantics at intermediate layers, these capabilities largely fail to surface in the final text generation when audio conflicts with vision. Probing analyses show that useful latent audio information is present, but deeper fusion layers disproportionately privilege visual representations that tend to suppress audio cues. We further trace this imbalance to training: the AVLLM's audio behavior strongly matches its vision-language base model, indicating limited additional alignment to audio supervision. Our findings reveal a fundamental modality bias in AVLLMs and provide new mechanistic insights into how multimodal LLMs integrate audio and vision.

1. Introduction

AVLLMs extend LLMs with audio and visual inputs, but their multimodal integration remains insufficiently understood and unreliable when modalities conflict. This study introduces a mechanistic analysis of how audio and visual representations evolve and influence generation.

Audio supplies off-screen events, speech, music, and ambient cues that vision alone cannot capture, motivating unified audio-visual models.
Interpretability has examined text-only, vision-language, and audio-language models, but audio-visual integration mechanisms remain unstudied.
In counterfactual scenes, AVLLMs hallucinate sounds associated with visible silent objects while ignoring the actual off-screen sound.
Audio understanding drops by up to 56% when audio and visual cues conflict, indicating heavy reliance on vision and underuse of audio.
The study analyzes attention across layers, probes audio representations, performs attention knockouts, and compares outputs with base vision-language models.
Intermediate layers retain interpretable audio semantics, but deeper layers prioritize visual features and suppress audio information during generation.

2. Related Work

Prior work provides foundations for adapter-based audio-visual modeling and mechanistic interpretability, but multimodal internal mechanisms remain comparatively underexplored. The paper builds on these architectural and analytical approaches.

Most AVLLMs connect frozen audio and visual encoders to a pretrained language backbone through learned adapters.
Early systems such as Video-LLaMA and PandaGPT use simple MLP adapters to map modality embeddings into the LLM token space.
Mechanistic interpretability has localized factual knowledge, capabilities, and processing patterns in text-only LLMs, while multimodal applications remain nascent.
Causal mediation with knockout techniques and probing with linear classifiers or logit lens provide tools for attributing component contributions and examining representations.

3. Preliminary

AVLLMs extend autoregressive transformer language models with projected visual and audio features, allowing multimodal inputs to be processed within a shared representation sequence. Their transformer layers progressively contextualize these embeddings through attention and feed-forward computations before producing next-token probabilities.

Transformer Language Models: The embedding layer maps each input token to a vector that initializes the sequence of representations refined by transformer layers.Each layer contains multi-head self-attention and feed-forward network sublayers with residual connections.
Transformer Language Models: Multi-head self-attention enables tokens to exchange information through weighted aggregations based on learned query-key similarities.The causal mask restricts each position to attending only to previous positions and itself.
Transformer Language Models: The final transformer representation is passed through an unembedding layer to produce logits over the vocabulary and a probability distribution for the next token.The unembedding uses matrix WU and bias b before softmax conversion.
Audio-Visual LLMs: AVLLMs combine pretrained vision and audio encoders with learned adapters that project modality features into the LLM representation space.The projected visual and audio embeddings are concatenated with text embeddings, and modality embeddings may be interleaved in practice.

4. Experimental Setup

The evaluation uses audio-visual captioning and counterfactual audio–video pairs to test whether models integrate both modalities. An LLM judge scores audio and video fidelity separately using detailed, calibrated reasoning.

Evaluation Task: The captioning task prompts models to “describe what you see and hear,” directly testing audio-visual perception and integration.Counterfactual samples swap the original audio track to isolate whether models process each modality independently or rely predominantly on vision.
Evaluation Task: The evaluation set contains 500 AudioCaps-derived videos, evenly divided between factual and counterfactual audio–video pairings.
Audio Visual Caption Evaluation: An LLM judge scores generated captions from 0 to 1 for audio and video fidelity after reasoning about objects, actions, temporal order, and audio events.Ratings are calibrated with few-shot in-context examples and accompanied by summarized reasoning.
Audio Visual Caption Evaluation: Human evaluation shows strong correlation between the judge and human ratings: Spearman ρ = 0.816 for audio fidelity and ρ = 0.732 for video fidelity.

5. Investigating Attention Pattern

Attention analysis reveals a layer-dependent modality asymmetry: audio receives substantial attention early, then nearly disappears, while video attention rises in deeper layers. Thus, models attend to audio initially but prioritize vision during later processing.

Attention Patterns: Audio tokens receive 40-50% attention in layers 0-5, but attention drops to near-zero in subsequent layers.
Attention Patterns: Video-token attention steadily increases through layers 15-30, reaching 20-40% as deeper layers process the input.
Attention Patterns: Query tokens dominate attention across all layers, capturing 60-100% despite comprising the fewest input tokens.
Attention Patterns: The resulting deeper-layer asymmetry gives vision substantial attention while audio is largely ignored.

6. Probing Audio Representations

Probing shows that intermediate audio representations encode meaningful sound semantics, including multilingual sound-source concepts. However, these latent representations coexist with much weaker manifested audio fidelity when visual cues conflict.

Probing Audio Representations: Logit-lens probing decodes hidden states at audio-token positions through the unembedding matrix to assess their vocabulary-level semantics.The method extracts the highest-logit token at each audio-token position for every layer.
Probing Audio Representations: Decoded representations meaningfully capture sound sources and event attributes, with sound-source concepts more consistent than actions or sound attributes.Examples of sources include drill, engine, horse, and keyboard; actions are less consistent and attributes are rare.
Probing Audio Representations: Audio representations map to sound-related concepts in multiple languages, including Chinese tokens for horse, keyboard, and door.The multilingual mapping is consistent with the Qwen base model’s substantial Chinese pretraining.
Latent Versus Manifested Audio: 61.4% latent audio understanding contrasts with 23% counterfactual audio caption fidelity, indicating that meaningful audio information is present internally but not incorporated reliably into generation.

7. Investigating Information Flow

The study traces audio-visual information flow with attention analysis and causal knockouts. Both modalities integrate in deeper layers, but visual pathways dominate final audio integration, while blocking vision can recover audio understanding.

Method: Attention knockouts block generated tokens from attending to audio or video at selected layers to identify each modality’s influence on text generation.The intervention modifies attention masks and measures changes in generated captions.
Factual Samples: In factual samples, blocking audio does not degrade video understanding, while blocking video causes a moderate ∼10% drop from middle layers.The model can compensate for blocked video by leveraging audio in aligned settings.
Counterfactual Samples: In counterfactual samples, blocking video produces severe video-understanding drops of ∼40% in final layers, revealing concentrated deep-layer visual integration.Mismatched audio and vision remove the complementary information available in factual samples.
Counterfactual Samples: Blocking audio in counterfactual samples loses up to ∼50% relative audio performance, showing that audio transfers primarily through deep layers.Audio and video therefore share a similar late integration pattern before their effects diverge.
Counterfactual Samples: Blocking visual pathways in final layers recovers approximately 50% relative audio performance, indicating that vision interferes with audio processing.The results identify visual dominance as the critical asymmetry in cross-modal integration.

8. Investigating Origins of Visual Bias

The paper investigates whether AVLLMs inherit visual bias from vision-language training. Token distributions remain highly similar to the vision-only base model, and qualitative examples show visually grounded audio hallucinations.

Method: AVLLMs are compared with their pretrained vision-language base models to test whether audio contributes meaningfully to generation.The analysis uses token-distribution similarity between Qwen2.5Omni and Qwen2.5VL.
Token Distribution Analysis: The average KL divergence is 0.4, indicating high similarity between AVLLM and base-LVLM output distributions despite audio input.This comparison tests whether alignment established balanced multimodal processing.
Token Distribution Analysis: 85.36% of audio-related tokens generated by the AVLLM are predictable from the top three choices of the vision-only base model.The token breakdown is 66.06% unshifted, 19.30% marginal, and 14.63% shifted.
Qualitative Examples: Qwen2.5Omni hallucinates helicopter sounds from visible video while the actual audio contains a crying baby and child speaking.The hallucinated sound tokens are unshifted and receive strong localized attention on the helicopter.
Qualitative Examples: Even when instructed to describe audio, the model can describe visible objects instead, and most predicted tokens remain unshifted relative to Qwen2.5VL.These examples support strong visual grounding in audio-related generation.

9. Conclusion, Limitations, and Future Work

The paper concludes that AVLLMs retain latent audio understanding but systematically favor vision when conflicting modalities must be transferred into text. It recommends counterfactual evaluation and training strategies that address modality imbalance, while limiting scope to open-source models and mainly nonspeech sounds.

Conclusion: Audio understanding severely degrades when audio and visual information conflict, despite meaningful audio representations in intermediate layers.The conclusion frames this as a systematic preference for visual cues during final generation.
Conclusion: Vision representations dominate deeper cross-modal transfer, while selectively blocking vision largely recovers latent audio understanding.This links the observed output failure to late-layer modality integration.
Future Work: The authors recommend counterfactual evaluation and balanced or counterfactual training data to penalize visual shortcuts.They also propose future methods to regularize modality bias within transformer layers.
Limitations: The analysis is limited to open-source AVLLMs and mainly examines nonspeech audio events.The paper identifies large-scale counterfactual-data curation and transformer-level regularization as future work.

Supplementary Material

The supplementary material describes factual and counterfactual evaluation data and the World Sense task. Counterfactual samples deliberately decouple audio from visible content to test independent modality processing.

A.1. Data Source: AudioCaps supplies videos with human-written captions describing their sounds, while visual descriptions are generated with GPT-4.1 and manually reviewed.AudioCaps provides five annotations per sample, and its scenes are generally simple and accurately captioned.
A.1. Data Source: Factual samples pair highly correlated visual and audio content, allowing audio events to potentially be inferred from vision.Figure 8 illustrates this aligned setting.
Additional Task: World Sense tasks couple perception and reasoning.The supplied figure identifies this as an example task.
Counterfactual Construction: Counterfactual samples pair videos with audio tracks that cannot plausibly be inferred from visible objects.They are designed to test whether AVLLMs process audio independently of vision.
Counterfactual Construction: The authors swap original audio tracks with semantically dissimilar tracks and select 250 pairs with lowest cosine similarity, approximately 0.498.The evaluation contains 250 factual and 250 counterfactual samples.

A.3. Existing Benchmarks

Existing audio-visual benchmarks do not adequately isolate perceptual modality bias because many combine perception with reasoning or fail to create sufficiently adversarial cross-modal conflicts.

Counterfactual dataset construction: The proposed counterfactual construction selects 250 audio-video pairs with approximately 0.498 semantic similarity, then synthesizes videos by swapping audio tracks.The pipeline computes audio and vision embeddings, performs Hungarian matching, filters candidate pairs, and creates counterfactual videos.
Benchmark limitations: Reasoning-dependent benchmarks conflate audio-visual perception with world-knowledge reasoning, obscuring isolated modality biases.The authors therefore seek tasks that separate perceptual capability from reasoning-dependent answering.
Benchmark limitations: 75.6% accuracy with both modalities versus 73.0% without video shows AVHBench’s visual cues fail to sufficiently mislead models in video-induced audio hallucination.The small difference indicates insufficient modality conflict for stress-testing independent integration.

B.1. Human Evaluation Study

The study validates its caption-fidelity judge against human ratings and extends mechanistic analyses across multiple AVLLMs. Across models, latent audio information exceeds generated audio fidelity, while deeper-layer visual pathways interfere with audio use and generation remains aligned with vision-language priors.

Human evaluation: Spearman’s ρ = 0.816 for audio fidelity and ρ = 0.732 for video fidelity indicates strong alignment between LLM-judge scores and human ratings.Two graduate annotators independently rated 200 stratified factual and counterfactual captions using the same rubric.
Cross-model generalization: MiniCPM-o2.6 reaches 75.4% latent audio recall versus 22.1% generated caption fidelity, while VideoLLaMA 2.1 reaches 59.9% versus 34.1%.These results generalize the competence gap between intermediate audio representations and final generated outputs beyond the Qwen-Omni series.
Cross-model generalization: Decoded intermediate audio tokens describe sound events such as sirens and barking, but multilingual audio representations appear specific to Qwen’s training data.MiniCPM-o2.6 and VideoLLaMA 2.1 lack the multilingual token representations observed in Qwen2.5-Omni.
Cross-modal information flow: Attention knockouts show that deeper layers process both modalities, yet blocking visual pathways can recover audio understanding, revealing visual interference.The replicated experiments support similar cross-modal integration patterns in MiniCPM-o2.6 and VideoLLaMA 2.1.
Origin of visual bias: InternOmni’s audio-visual output distribution has KL divergence 0.46 from InternVL’s vision-only distribution, with 70.62% of audio-related tokens unshifted.The alignment suggests generation remains dominated by the vision-language base model’s priors and that modality imbalance may be inherited from initialization and alignment.