Towards Mitigating Hallucinations in Large Vision-Language Models by Refining Textual Embeddings
Aakriti Agrawal, Gouthaman KV, Rohith Aralikatti, Gauri Jagatap, Jiaxin Yuan, Vijay Kamarshi, Andrea Fanelli, Furong Huang
2025-11-10
Summary
This research focuses on a problem with how current visual language models, which process both images and text, tend to rely too much on the text information and not enough on the visual information.
What's the problem?
Many visual language models are built by simply adding information about the image to the text input. This creates a bias where the model prioritizes understanding the text and often ignores or misinterprets the image, leading to incorrect answers or even 'hallucinations' where the model makes things up that aren't actually in the image.
What's the solution?
The researchers found a simple way to fix this by improving the text understanding *using* information from the image. They took the key features from the image, averaged them together, and then used that average to refine how the model understands the text. This helps the model pay more attention to what's actually in the image.
Why it matters?
This work is important because it identifies a fundamental flaw in how these models are designed and offers a straightforward solution that makes them more accurate and reliable. By reducing hallucinations, the models become more trustworthy for tasks like image captioning or answering questions about images, and it sets the stage for even more advanced ways to combine visual and textual information in the future.
Abstract
In this work, we identify an inherent bias in prevailing LVLM architectures toward the language modality, largely resulting from the common practice of simply appending visual embeddings to the input text sequence. To address this, we propose a simple yet effective method that refines textual embeddings by integrating average-pooled visual features. Our approach demonstrably improves visual grounding and significantly reduces hallucinations on established benchmarks. While average pooling offers a straightforward, robust, and efficient means of incorporating visual information, we believe that more sophisticated fusion methods could further enhance visual grounding and cross-modal alignment. Given that the primary focus of this work is to highlight the modality imbalance and its impact on hallucinations -- and to show that refining textual embeddings with visual information mitigates this issue -- we leave exploration of advanced fusion strategies for future work.