AVCD: Mitigating Hallucinations in Audio-Visual Large Language Models through Contrastive Decoding

TL;DR

AVCD reduces hallucinations in audio-visual LLMs via contrastive decoding, improving accuracy on AVHBench dataset.

cs.CV 🔴 Advanced 2025-05-27 3 views
Chaeyoung Jung Youngjoon Jang Joon Son Chung
multimodal contrastive decoding hallucination audio-visual large language model

Key Findings

Methodology

AVCD is a novel decoding framework that uses attention distributions to dynamically identify less dominant modalities and applies attentive masking to generate perturbed output logits. It reformulates the contrastive decoding framework to support audio, visual, and textual inputs in a trimodal setting.

Key Results

  • On the AVHBench dataset, AVCD improves VideoLLaMA2's accuracy by 2% and video-SALMONN by 7%.
  • On the MUSIC-AVQA dataset, AVCD achieves an accuracy of 81.58%, significantly outperforming other methods.
  • Entropy-guided adaptive decoding in AVCD reduces computational overhead while maintaining accuracy.

Significance

AVCD addresses hallucination issues in multimodal large language models, particularly in audio-visual scenarios. It provides a more reliable decoding strategy for multimodal interactions, enhancing model performance in complex tasks.

Technical Contribution

AVCD introduces dynamic modality recognition and attentive masking strategies, offering a more flexible decoding scheme compared to existing fixed modality perturbation methods. It also improves computational efficiency through entropy-guided adaptive decoding.

Novelty

AVCD is the first to apply contrastive decoding in audio-visual LLMs, dynamically identifying modality dominance and applying masking to significantly reduce hallucinations.

Limitations

  • AVCD may perform poorly in extreme noise environments as it relies on the accuracy of attention distributions.
  • Entropy-guided adaptive decoding may still face performance bottlenecks under high computational loads.

Future Work

Future work could explore AVCD's application in more modality combinations, such as incorporating tactile or other sensor data. Further optimization of adaptive decoding strategies to enhance efficiency is also a key direction.

AI Executive Summary

Multimodal large language models (MLLMs) have shown great potential in handling complex tasks, but hallucinations remain a major challenge. Existing contrastive decoding methods have achieved some success in vision-language models but are limited in audio-visual LLMs, where hallucinations often arise from unimodal and cross-modal combinations of audio, video, and language. To address this, the paper proposes a novel decoding framework—Audio-Visual Contrastive Decoding (AVCD).

AVCD dynamically identifies less dominant modalities and applies attentive masking to generate perturbed output logits, suppressing modality-induced hallucinations. Unlike previous methods that perturb a fixed modality, AVCD leverages attention distributions to identify modality dominance and improves efficiency through entropy-guided adaptive decoding. Experimental results show that AVCD consistently outperforms existing methods across multiple datasets, particularly improving accuracy by 2% for VideoLLaMA2 and 7% for video-SALMONN on the AVHBench dataset.

Despite its effectiveness in reducing hallucinations, AVCD may face challenges in extreme noise environments. Future research could explore its application in more modality combinations and further optimize decoding strategies to enhance efficiency. AVCD provides a more reliable decoding scheme for multimodal interactions, with significant academic and industrial implications.

Deep Analysis

Background

Multimodal large language models (MLLMs) combine visual and auditory inputs to enhance the model's ability to understand and solve complex tasks. However, hallucinations remain a major challenge, especially in scenarios involving audio-visual combinations. Existing contrastive decoding methods have achieved some success in vision-language models but are limited in audio-visual LLMs.

Core Problem

Hallucination refers to the generation of biased or factually incorrect information, which is particularly prominent in multimodal large language models. The hallucination problem in audio-visual LLMs is especially complex as it involves unimodal and cross-modal combinations of audio, video, and language.

Innovation

AVCD dynamically identifies less dominant modalities and applies attentive masking to generate perturbed output logits, suppressing modality-induced hallucinations. Unlike previous methods that perturb a fixed modality, AVCD leverages attention distributions to identify modality dominance and improves efficiency through entropy-guided adaptive decoding.

Methodology

  • �� Dynamic Modality Recognition: Uses attention distributions to identify less dominant modalities.
  • �� Attentive Masking: Applies masking to less dominant modalities to generate perturbed output logits.
  • �� Entropy-Guided Adaptive Decoding: Selectively skips unnecessary decoding steps based on model confidence.

Experiments

Experiments were conducted on multiple datasets, including MUSIC-AVQA and AVHBench. The effectiveness of AVCD in reducing hallucinations was validated through comparisons with existing methods. The experiments also evaluated the efficiency gains from entropy-guided adaptive decoding.

Results

AVCD improves VideoLLaMA2's accuracy by 2% and video-SALMONN by 7% on the AVHBench dataset. On the MUSIC-AVQA dataset, AVCD achieves an accuracy of 81.58%, significantly outperforming other methods.

Applications

AVCD can be applied to complex tasks requiring multimodal interactions, such as audio-visual data analysis in autonomous driving and multimodal human-computer interaction in smart homes.

Limitations & Outlook

AVCD may perform poorly in extreme noise environments as it relies on the accuracy of attention distributions. Entropy-guided adaptive decoding may still face performance bottlenecks under high computational loads.

Plain Language Accessible to non-experts

Imagine a kitchen where a chef needs to handle multiple tasks: chopping vegetables, boiling soup, and baking bread. Each task requires different tools and attention. AVCD acts like a smart kitchen assistant, identifying which task needs the most attention and helping the chef allocate focus, ensuring each dish is perfectly prepared. In this way, AVCD reduces errors caused by over-relying on one task, just like avoiding burning the bread by forgetting to turn off the oven.

ELI14 Explained like you're 14

Imagine you're playing a super complex game where you control multiple characters. Each character has its own task, like one is fighting monsters, another is collecting resources, and another is solving puzzles. AVCD is like a super smart game assistant that tells you which character needs your attention the most, so you don't mess up by focusing on just one character. Isn't that cool?

Glossary

Contrastive Decoding

A method that reduces hallucinations by contrasting original logits with perturbed logits.

Used to reduce hallucinations in multimodal large language models.

Attention Distribution

The focus level of the model on different modalities during decoding.

Used to identify less dominant modalities.

Entropy-Guided Adaptive Decoding

Selectively skips unnecessary decoding steps based on model confidence.

Improves decoding efficiency.

Modality

Refers to different types of data inputs like audio, video, or text.

Types of inputs involved in multimodal large language models.

Hallucination

The generation of biased or factually incorrect information by the model.

A problem to be addressed in multimodal large language models.

Open Questions Unanswered questions from this research

  • 1 How to improve AVCD's performance in extreme noise environments? Current methods rely on the accuracy of attention distributions, requiring more robust strategies.
  • 2 How to further optimize entropy-guided adaptive decoding strategies to enhance computational efficiency?

Applications

Immediate Applications

Autonomous Driving

AVCD can be used to analyze audio-visual data in autonomous driving, improving decision accuracy and safety.

Smart Homes

In smart homes, AVCD can enhance multimodal human-computer interaction, improving user experience.

Long-term Vision

Multimodal AI Systems

The application of AVCD can drive the development of multimodal AI systems, enabling more complex task processing and decision-making.

Abstract

Hallucination remains a major challenge in multimodal large language models (MLLMs). To address this, various contrastive decoding (CD) methods have been proposed that contrasts original logits with hallucinated logits generated from perturbed inputs. While CD has shown promise in vision-language models (VLMs), it is not well-suited for AV-LLMs, where hallucinations often emerge from both unimodal and cross-modal combinations involving audio, video, and language. These intricate interactions call for a more adaptive and modality-aware decoding strategy. In this paper, we propose Audio-Visual Contrastive Decoding (AVCD)-a novel, training-free decoding framework designed to model trimodal interactions and suppress modality-induced hallucinations in AV-LLMs. Unlike previous CD methods in VLMs that corrupt a fixed modality, AVCD leverages attention distributions to dynamically identify less dominant modalities and applies attentive masking to generate perturbed output logits. To support CD in a trimodal setting, we also reformulate the original CD framework to jointly handle audio, visual, and textual inputs. Finally, to improve efficiency, we introduce entropy-guided adaptive decoding, which selectively skips unnecessary decoding steps based on the model's confidence in its predictions. Extensive experiments demonstrate that AVCD consistently outperforms existing decoding methods. Especially, on the AVHBench dataset, it improves accuracy by 2% for VideoLLaMA2 and 7% for video-SALMONN, demonstrating strong robustness and generalizability. Our code is available at https://github.com/kaistmm/AVCD.

cs.CV