Mitigating Object Hallucinations in LVLMs via Attention Imbalance Rectification
Researchers are developing new methods to combat hallucinations in AI models, particularly in multimodal systems. One approach focuses on retrieval-augmented reliability-aware inference, which uses an external database to estimate prediction trustworthiness and abstain from answering when evidence is insufficient. Another method addresses semantic hallucination in explainable AI for vision-language models by disentangling unique semantic signals. Additionally, a technique called Attention Imbalance Rectification aims to reduce object hallucinations in Large Vision-Language Models by adjusting attention allocation. Finally, a study reformulates token-level hallucination detection as a quickest change detection problem to improve reaction time. AI
IMPACT These research papers introduce novel techniques to improve the reliability and trustworthiness of AI models by reducing hallucinations, which is crucial for their deployment in sensitive applications.
- Lorden's lower bound
- Cusumano
- Donsker-Varadhan
- Quickest Detection of Hallucination Onset: Delay Bounds and Learned CUSUM Statistics
- Multimodal Large Language Models
- alphaXiv
- ScienceCast
- CatalyzeX
- ImageNet-100
- Gotit.pub
- MM-Vet
- POPE
- Linear Semantic Attribution
- Hugging Face
- Attention Imbalance Rectification
- arXiv
- Vision-Language Models
- LVLMs
- DagsHub