Gemma-3

Gemma-3 is a dataset/benchmark used to evaluate hallucination detection in vision-language models by probing their internal representations before any text generation occurs.

Papers using Gemma-3 (3)

Cascaded Sparse Autoencoders Learn Multi-Level Visual Concepts in Multimodal LLMs2026HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single Token2026MASSV: Multimodal Adaptation and Self-Data Distillation for Speculative Decoding of Vision-Language Models2025