Mitigating Multimodal LLMs Hallucinations via Relevance Propagation at Inference Time
Abstract
Multimodal large language models (MLLMs) have revolutionized the landscape of AI, demonstrating impressive capabilities in tackling complex vision and audio-language tasks. However, a critical challenge remains: these models often suffer from hallucinations, generating outputs that diverge from the provided perceptual inputs. This tendency stems from an inherent imbalance in modality utilization during inference, where the dominance of textual tokens undermines the potential of perceptual inputs. As a result, the model frequently resorts to textual language priors at the expense of grounded evidence. To tackle this issue, we propose Learning Inference-time Modality Enhancement (LIME), a training-free framework designed to bolster multimodal grounding by explicitly enhancing modality usage during decoding. LIME leverages Layer-wise Relevance Propagation (LRP) to quantify token-level contributions and defines a relevance-based objective that promotes increased reliance on perceptual inputs. This objective is enforced through inference-time updates to the model's key-value representations, without modifying model parameters or requiring additional training data. We evaluate LIME across multiple multimodal benchmarks in both vision and audio domains, demonstrating consistent reductions in hallucinations and enhanced grounding while preserving generation quality. Further analysis shows that LIME increases modality contribution and produces more localized and semantically aligned relevance patterns.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
This paper presents a novel approach to enhancing modality utilization in multimodal large language models, significantly improving their reliability and grounding capabilities. The methodology and experimental results indicate a meaningful contribution to the field, addressing a critical challenge in the deployment of multimodal AI systems.
Methodology
The proposed method, Learning Inference-time Modality Enhancement (LIME), is innovative in its approach to mitigating hallucinations in multimodal large language models (MLLMs) by leveraging Layer-wise Relevance Propagation (LRP) to adjust key-value representations during inference. This training-free framework enhances modality utilization without altering model parameters, which is a significant advancement over existing methods that typically require retraining or fine-tuning. The methodology is well-structured, with a clear explanation of how relevance scores are computed and utilized to optimize the model's performance dynamically during decoding.
Results & evaluation
The experiments are comprehensive, covering multiple benchmarks in both vision and audio domains, demonstrating the effectiveness of LIME in reducing hallucinations while maintaining generation quality. The results show consistent improvements across various metrics, indicating that the method is robust and generalizable. However, the paper could benefit from more extensive ablation studies to isolate the effects of different components of the approach.
Reproducibility
The paper provides sufficient details regarding the implementation of LIME, including hyperparameters and optimization procedures. The source code is publicly available, which enhances reproducibility. However, the absence of specific dataset splits and evaluation protocols in the main text may hinder complete reproducibility for some readers.
Limitations
While LIME shows promise, it incurs additional computational overhead due to the optimization steps required at each decoding stage, which may limit its applicability in real-time or resource-constrained environments. Additionally, the method's reliance on LRP may introduce challenges in interpretability and stability in certain contexts.
Broader impact
The findings of this research have significant implications for the reliability of multimodal AI systems, particularly in applications where accurate grounding in perceptual data is critical, such as autonomous systems, assistive technologies, and interactive AI. By improving the grounding of multimodal models, this work contributes to the development of more trustworthy AI systems that can better understand and respond to complex inputs. This paper presents a novel approach to enhancing modality utilization in multimodal large language models, significantly improving their reliability and grounding capabilities. The methodology and experimental results indicate a meaningful contribution to the field, addressing a critical challenge in the deployment of multimodal AI systems.
Author of this paper? Send a correction.