Get Started
Home
Topics
Search
Library
Research questionHow can we detect object hallucinations in vision-language models when visual grounding shifts across layers?A model may mention objects that are absent from an image, and its visual grounding can change substantially across internal layers. Reliable detection therefore requires distinguishing predictions supported by visual evidence from hallucinated ones.
AI
Alignment & Safety
Computer Vision
Evaluation & Benchmarks
Multimodal Models
Latest papersRecent research connected to this question, newest first.Detecting Object Hallucinations in Large Vision-Language Models via Cross-Modal Attention Drifts and Mask-Based VerificationApplies to large vision-language models during decoding and requires access to cross-modal attention information, prediction probabilities, and the ability to test predictions with targeted image masking. The reported evidence covers multiple benchmarks and representative open-source LVLMs; applicability to models without these accessible signals is not established.research paper · Sep 2, 2026
Related questions
How can we reliably detect when an LLM response is unsupported by its reference documents?How can large vision-language models reduce visual hallucinations without extra training or decoding passes?How can grounded language-model systems preserve object identity from selection through evidence delivery?How can hallucination detectors verify claims and localize unsupported spans without costly separate pipelines?