Truth Finding in Image-Text AI: The TruthPrInt Method Against Hallucinations

Large Vision-Language Models (LVLMs) have made impressive progress in recent years. They can describe images, answer questions about visual scenes, and even generate stories related to images. Despite this progress, LVLMs struggle with a persistent problem: object hallucinations (OH). This means the models invent objects that are not actually present in the image. This poses a significant obstacle for the use of LVLMs in critical applications, such as medical image analysis or autonomous navigation.

A new research approach called TruthPrInt (Truthful-Guided Pre-Intervention) promises to address this problem. The method is based on the realization that the internal states of LVLMs, particularly the so-called hidden states, contain information about the "truthfulness" of the generated statements. However, until now, it was unclear how these internal states function in the context of object hallucinations and whether they can serve as indicators for hallucinations at the token level.

The researchers behind TruthPrInt have thoroughly investigated the internal states of LVLMs and made two important discoveries. First, the internal states are highly specific indicators of hallucinations at the token level. Second, different LVLMs encode universal patterns of hallucinations in shared latent subspaces. This suggests that there are "generic truth directions" shared by different LVLMs.

Based on these findings, TruthPrInt works in two steps. First, the model learns the "truth direction" of the LVLM decoding. In the second step, a truth-guided intervention is performed during decoding. This intervention steers the model's generation in a way that reduces the probability of hallucinations.

To improve the transferability of hallucination detection between different LVLMs and datasets, the researchers also developed ComnHallu. ComnHallu constructs and aligns hallucination subspaces to identify and utilize common patterns.

TruthPrInt has been evaluated in extensive experiments, both in in-domain and out-of-domain scenarios, with common LVLMs and OH benchmarks. The results show that TruthPrInt significantly outperforms the current state-of-the-art. The method thus offers a promising approach to improving the reliability of LVLMs and paves the way for their use in demanding applications.

The reduction of hallucinations is a crucial step in strengthening trust in AI systems. Methods like TruthPrInt contribute to improving the robustness and reliability of LVLMs and open up new possibilities for their use in various fields.

Bibliographie: - https://arxiv.org/abs/2503.10602 - https://arxiv.org/html/2503.10602v1 - https://www.researchgate.net/publication/384234767_Mitigating_Object_Hallucinations_in_Large_Vision-Language_Models_through_Visual_Contrastive_Decoding - https://aclanthology.org/2024.findings-emnlp.685.pdf - https://dl.acm.org/doi/10.1145/3703155 - https://openreview.net/forum?id=eFoj2egr7G - https://openreview.net/forum?id=kX2EqGUp5B&referrer=%5Bthe%20profile%20of%20Ziniu%20Li%5D(%2Fprofile%3Fid%3D~Ziniu_Li1) - https://github.com/NishilBalar/Awesome-LVLM-Hallucination - https://openaccess.thecvf.com/content/CVPR2024/papers/Leng_Mitigating_Object_Hallucinations_in_Large_Vision-Language_Models_through_Visual_Contrastive_CVPR_2024_paper.pdf - https://github.com/showlab/Awesome-MLLM-Hallucination/blob/main/README.md