[arXiv]score: 0.23

GHOST-Q: Quantization Increases Hallucinations in 8B Vision-Language Models

September 25, 2026

GHOST-Q reveals that while quantization (INT8, NF4) preserves MMStar accuracy within ±2% for 8B VLMs, it significantly redistributes grounding failures. Quantization often preserves headline accuracy scores while increasing hallucination rates in specific visual grounding conditions.

HOW THIS AFFECTS YOU

●

builderYou should not rely on aggregate accuracy benchmarks alone when deploying quantized VLMs due to increased grounding hallucinations.

●

researcherYou can use item-by-item paired prediction analysis to detect how compression affects specific failure modes.

read original ↗arxiv.org

DAILY DIGEST

you don't check 9 sources — we do. one email every morning, read in 2 min. free. unsubscribe anytime. privacy

← back to feed