Selective Prior Calibration Reduces Commonsense-Driven Hallucinations in Vision-Language Models
A new research paper on arXiv (2607.29240) addresses commonsense-driven hallucination (CDH) in vision-language models, where a model's prior knowledge overrides clear visual evidence of atypical states. The authors propose Selective Prior Calibration (SPC), a method that selectively adjusts candidate-level prior-preference estimates based on instance-dependent strength, revising predictions only when the score pattern strongly supports an alternative. Experiments demonstrate the method's effectiveness in mitigating CDH while preserving correct answers on commonsense images. The paper is relevant to AI technology and digital art applications, as it improves the reliability of AI systems in interpreting visual content.
Key facts
- Paper arXiv:2607.29240 introduces Selective Prior Calibration (SPC) to mitigate commonsense-driven hallucination in vision-language models.
- CDH occurs when a model's commonsense prior overrides visual evidence, e.g., reporting a six-fingered hand as five-fingered.
- SPC subtracts candidate-level prior-preference estimates from image-conditioned scores with instance-dependent strength.
- SPC revises predictions only when the score pattern strongly supports an alternative.
- Experiments show SPC repairs counterfactual errors while preserving correct answers on matched commonsense images.
- The paper is available on arXiv under the identifier 2607.29240.
- The research is relevant to AI technology and digital art, improving AI interpretation of visual content.
Entities
Institutions
- arXiv