MoD-DPO: Towards Mitigating Cross-modal Hallucinations in Omni LLMs using Modality Decoupled Preference Optimization
MoD-DPO reduces cross-modal hallucinations via modality decoupling, boosting accuracy to 88.19%.
Key Findings
Methodology
MoD-DPO employs a modality-decoupled direct preference optimization framework, introducing modality-aware regularization terms to ensure model invariance to irrelevant modality perturbations and sensitivity to relevant ones, thus reducing unintended cross-modal interactions. Additionally, a language-prior debiasing penalty reduces hallucination-prone text-only responses.
Key Results
- MoD-DPO improved accuracy on AVHBench to 88.19%, a 3% increase over OmniDPO.
- On the CMM benchmark, MoD-DPO++ showed significant improvement in language dominance tasks, demonstrating the effectiveness of language-prior debiasing.
- Ablation studies confirmed the critical role of modality decoupling and language-prior debiasing in enhancing model performance.
Significance
This research significantly enhances the reliability and robustness of multimodal foundation models by reducing cross-modal hallucinations. It addresses long-standing issues of hallucinations caused by language bias and spurious inter-modality correlations, offering new insights for building more reliable multimodal systems.
Technical Contribution
MoD-DPO introduces modality-decoupled regularization terms and language-prior debiasing mechanisms, distinguishing it from existing preference optimization methods. It provides stronger theoretical guarantees for modality alignment and demonstrates more efficient training paths for multimodal models.
Novelty
MoD-DPO is the first to incorporate modality decoupling into preference optimization, significantly reducing cross-modal hallucinations. Unlike existing methods, it innovatively addresses spurious inter-modality correlations through explicit modality-aware regularization.
Limitations
- In complex multimodal scenarios, the model may still experience hallucinations, especially when inter-modality correlations are weak.
- Training on large datasets is time-consuming and computationally expensive.
Future Work
Future research could explore more efficient modality decoupling methods to further reduce computational costs. Additionally, applying this method to a wider range of multimodal tasks could be investigated.
AI Executive Summary
Recent advances in multimodal large language models have achieved remarkable progress in audiovisual understanding tasks. However, these models remain susceptible to cross-modal hallucinations, often arising from spurious correlations and language biases. In this context, the paper proposes a novel preference optimization framework, MoD-DPO, which significantly reduces cross-modal hallucinations through modality decoupling and language-prior debiasing.
MoD-DPO introduces modality-aware regularization terms to ensure model invariance to irrelevant modality perturbations and sensitivity to relevant ones, thus reducing unintended cross-modal interactions. Additionally, a language-prior debiasing penalty reduces hallucination-prone text-only responses. Experimental results demonstrate that MoD-DPO excels across multiple audiovisual hallucination benchmarks, with significant accuracy improvements.
This research not only provides stronger theoretical guarantees for modality alignment but also demonstrates more efficient training paths for multimodal models. Future research could explore more efficient modality decoupling methods to further reduce computational costs and apply this method to a wider range of multimodal tasks.
Deep Analysis
Background
In recent years, multimodal large language models have made significant progress in audiovisual understanding tasks by integrating visual, audio, and textual information. These models have achieved strong cross-modal understanding capabilities. However, due to spurious inter-modality correlations and language biases, they remain prone to cross-modal hallucinations, affecting their reliability and robustness.
Core Problem
Cross-modal hallucinations are a core issue in multimodal models, often stemming from spurious inter-modality correlations and language biases. These hallucinations not only affect model accuracy but can also lead to incorrect decisions and inferences. Therefore, effectively reducing cross-modal hallucinations is a major challenge in current multimodal model research.
Innovation
MoD-DPO introduces a modality-decoupled preference optimization framework, innovatively addressing cross-modal hallucinations. The method uses modality-aware regularization terms to ensure model invariance to irrelevant modality perturbations and sensitivity to relevant ones. Additionally, a language-prior debiasing penalty reduces hallucination-prone text-only responses.
Methodology
- �� Introduce modality-aware regularization terms to ensure model invariance to irrelevant modality perturbations.
- �� Implement a language-prior debiasing penalty to reduce hallucination-prone text-only responses.
- �� Utilize an automatically generated preference dataset for optimization, ensuring model stability across different modalities.
Experiments
Experiments were conducted on multiple audiovisual hallucination benchmarks, including AVHBench and CMM. The benchmarks included various audiovisual tasks to evaluate the model's cross-modal hallucination reduction effectiveness. Results showed that MoD-DPO performed exceptionally well across multiple benchmarks, with significant accuracy improvements.
Results
MoD-DPO improved accuracy on AVHBench to 88.19%, a 3% increase over OmniDPO. On the CMM benchmark, MoD-DPO++ showed significant improvement in language dominance tasks, demonstrating the effectiveness of language-prior debiasing.
Applications
The method can be directly applied to multimodal information processing systems, such as intelligent assistants, autonomous driving, and multimedia content analysis. By reducing cross-modal hallucinations, it enhances system reliability and robustness.
Limitations & Outlook
Despite MoD-DPO's excellent performance in reducing cross-modal hallucinations, the model may still experience hallucinations in complex multimodal scenarios. Additionally, training on large datasets is time-consuming and computationally expensive.
Plain Language Accessible to non-experts
Imagine you're in a kitchen cooking, with many tools and ingredients around. You need to choose the right tool for each ingredient, like using a knife to chop vegetables and a pan to fry them. MoD-DPO is like a smart chef who can pick the right tool for each ingredient, avoiding mistakes from using the wrong tool. This way, it reduces unnecessary errors and improves cooking efficiency and accuracy.
ELI14 Explained like you're 14
Hey there! Imagine you're playing a super cool game that needs you to use your eyes, ears, and hands all at once. Sometimes, the game might give you mixed-up info, like hearing sounds that aren't there or seeing things that don't exist. MoD-DPO is like a super helper that filters out this mixed-up info, letting you focus on what's really important. This way, you can finish game tasks faster and score higher!
Glossary
Cross-modal Hallucination
Refers to errors in perception by multimodal models due to spurious inter-modality correlations or language biases.
In this paper, cross-modal hallucination is the main research focus, and MoD-DPO aims to reduce this phenomenon.
Modality Decoupling
Separating information from different modalities to reduce interference and spurious correlations.
MoD-DPO reduces cross-modal hallucinations through modality decoupling.
Language Bias
A phenomenon where multimodal models overly rely on language modality, ignoring other modality information.
MoD-DPO reduces the impact of language bias through a debiasing penalty.
Preference Optimization
Optimizing model preference selection to improve performance on specific tasks.
MoD-DPO uses a preference optimization framework to enhance model accuracy.
Regularization Term
Constraints added during optimization to prevent model overfitting.
MoD-DPO introduces modality-aware regularization terms to ensure model stability.
Open Questions Unanswered questions from this research
- 1 How can MoD-DPO be efficiently applied to larger datasets?
- 2 Are there better methods to further reduce the impact of language bias on multimodal models?
Applications
Immediate Applications
Intelligent Assistants
By reducing cross-modal hallucinations, it improves the accuracy and reliability of intelligent assistants in multimodal information processing.
Long-term Vision
Autonomous Driving
Applying MoD-DPO in autonomous driving systems reduces perception errors during sensor data fusion, enhancing driving safety.
Abstract
Omni-modal large language models (omni LLMs) have recently achieved strong performance across audiovisual understanding tasks, yet they remain highly susceptible to cross-modal hallucinations arising from spurious correlations and dominant language priors. In this work, we propose Modality-Decoupled Direct Preference Optimization (MoD-DPO), a simple and effective framework for improving modality grounding in omni LLMs. MoD-DPO introduces modality-aware regularization terms that explicitly enforce invariance to corruptions in irrelevant modalities and sensitivity to perturbations in relevant modalities, thereby reducing unintended cross-modal interactions. To further mitigate over-reliance on textual priors, we incorporate a language-prior debiasing penalty that discourages hallucination-prone text-only responses. Extensive experiments across multiple audiovisual hallucination benchmarks demonstrate that MoD-DPO consistently improves perception accuracy and hallucination resistance, outperforming previous preference optimization baselines under similar training budgets. Our findings underscore the importance of modality-faithful alignment and demonstrate a scalable path toward more reliable and resilient multimodal foundation models.