MultiToP: Learning to Patch Visual Tokens to Mitigate Hallucinations in Video Large Multimodal Models

Problem
Video Large Multimodal Models (VLMMs) have shown significant advancements in video understanding tasks; however, they are still susceptible to hallucinations—instances where the generated outputs do not accurately reflect the input video content. This paper addresses the gap in existing literature regarding the mitigation of such hallucinations, proposing a novel approach to enhance the reliability of visual tokens used in these models. Notably, this work is a preprint and has not undergone peer review.

Method
The authors present MultiToP, a multimodal-context-aware visual token patching framework designed to refine unreliable visual tokens prior to language generation. The core technical contribution is the Visual Token Patcher, which predicts token-level replacement distributions to selectively substitute unreliable visual tokens with a dynamic global patch token. The training of the patcher employs information-guided rank calibration, leveraging answer-conditioned frame-level information cues from the backbone model to inform token replacement decisions. This method is augmented with ground-truth answer supervision and sparsity regularization, allowing for localized visual evidence refinement without necessitating modifications to the original model architecture.

Results
Extensive experiments demonstrate that MultiToP significantly reduces hallucinations on the Vript-HAL benchmark, achieving an impressive F1 score improvement of 50.60% over the baseline Qwen3-VL-4B-Instruct model. Additionally, MultiToP maintains the general video understanding capabilities of the model, yielding an 18.58% relative accuracy gain on the ActivityNet-QA benchmark when evaluated with Video-LLaVA-7B. These results indicate that the proposed framework not only enhances the fidelity of generated outputs but also preserves the model’s overall performance in video comprehension tasks.

Limitations
The authors acknowledge that while MultiToP effectively reduces hallucinations, the approach may introduce additional complexity in the training process due to the reliance on answer-conditioned cues. Furthermore, the framework’s performance may vary across different types of video content and tasks, which warrants further investigation. The paper does not address potential scalability issues when applied to larger models or datasets, nor does it explore the generalizability of the method across diverse multimodal applications.

Why it matters
The implications of this work are significant for the development of more reliable VLMMs, particularly in applications where accurate video understanding is critical, such as automated content generation and interactive AI systems. By addressing the hallucination problem, MultiToP paves the way for future research focused on enhancing the robustness of multimodal models. This advancement is crucial for building trust in AI systems that rely on video data, as highlighted in the findings available on arXiv.