Look Clearly Before Answering: Mitigating Hallucinations in LVLMs via Saliency-Driven Perceptual Realignment

In the rapidly evolving world of artificial intelligence, a new generation of systems has emerged that can see and speak at the same time. These large vision-language models act as digital observers, capable of looking at a photograph and describing what is happening within it, or answering questions about the scene. They have learned to connect the visual world with human language, creating a bridge between pixels and words. However, despite their impressive ability to generate fluent and often sensible sentences, these systems suffer from a persistent and frustrating flaw: they frequently lie. They might confidently describe a person holding a red balloon when the image shows a blue kite, or invent details that simply do not exist. This tendency to generate false information, known as hallucination, stems from a deep reliance on what the model has read before rather than what it is actually seeing right now. When the visual evidence is unclear or complex, the system often defaults to guessing based on its internal library of language patterns, producing answers that sound correct but are factually wrong. This unreliability prevents these powerful tools from being trusted in real-world situations where accuracy is essential.

Researchers at Xidian University and Tsinghua University have identified the root causes of this visual confusion and developed a method to help these models "look clearly" before they answer. Their work reveals that the problem is not just a lack of knowledge, but a failure in how the model processes and remembers what it sees during the split second it is thinking. The team discovered that as the model analyzes an image, its attention is often hijacked by unimportant background details, causing it to ignore the most critical parts of the picture. Furthermore, as the model begins to generate its response word by word, its memory of the visual scene degrades, becoming distorted and less connected to the specific question being asked. To fix this, the researchers introduced a training-free framework called Saliency-Driven Perceptual Realignment. This approach does not require retraining the massive AI model from scratch; instead, it acts as a guide during the thinking process, gently correcting the model's focus and memory to ensure it stays grounded in the visual reality of the image.

The first step in this correction process addresses the moment the model first looks at the image. The researchers found that the model's internal attention mechanism often gets stuck on "sink tokens," which are essentially meaningless background elements that accidentally grab too much focus. Imagine a person trying to listen to a conversation in a noisy room; if their attention is hijacked by a loud, irrelevant noise, they miss the important words being spoken. Similarly, the model was letting these background distractions drown out the salient visual cues, such as a person's face or a specific object. The new method detects when this hijacking occurs and redistributes the model's attention, pulling focus away from the noise and back toward the meaningful parts of the image. By doing this, the model recovers the suppressed visual evidence it was previously ignoring, allowing it to see the scene more accurately before it even begins to speak.

Once the model has a clearer view, the second challenge arises as it starts to generate its answer. The model stores a memory of the image to refer back to as it writes each new word. However, the researchers observed that this memory becomes distorted over time. Because the model processes the image and the question in a specific order, the memory of the visual details becomes tied to their position in the sequence rather than their actual importance to the question. It is as if the model forgets that a specific object is important because it was seen early in the sequence, while later, less important details become overly prominent in its memory. To counteract this, the new framework realigns this internal memory. It injects a signal that highlights which parts of the image are actually relevant to the current question, ensuring that the model's memory remains focused on the right visual evidence throughout the entire generation process, rather than drifting toward irrelevant details.

Finally, even with a clear view and a good memory, the model still faces the temptation to rely on its pre-existing language habits. Sometimes, the model knows the answer is "yes" based on the picture, but its internal language patterns strongly suggest "no" because that is a common phrase in its training data. To prevent this, the researchers added a final check that compares the model's visual-based prediction against a reference built from its own language habits. If the two disagree, the system actively suppresses the language-based guess and forces the model to stick with the visual evidence. This contrastive process ensures that the final answer is driven by what is actually in the image, not by what the model expects to see.

The results of applying this three-part strategy are significant. When tested across various different large vision-language models, the method consistently reduced the number of hallucinations and improved the accuracy of the answers. In specific tests measuring how often models correctly identify objects, the new approach showed improvements of up to nearly seven points, and in tests measuring the frequency of hallucinated details, it reduced errors by more than a third. Crucially, all of this was achieved without the need for expensive retraining or additional data, simply by adjusting how the model pays attention and remembers information while it works. By systematically addressing the degradation of visual awareness from the initial glance to the final sentence, this research offers a robust way to make these powerful AI systems more reliable, ensuring they tell the truth about what they see.