RESEARCHInvestigateNEXT 12 MONTHS
Context-Dependent Affordance Computation in Vision-Language Models
arXiv cs.CL — Computation and Language
Factual evidence
What the source reports
Research finds Vision-Language Models (Qwen3-VL-30B-A3B, LLaVA-1.5-13B) exhibit significant context-dependent affordance drift.
OneBench interpretation
Institutional assessment
So what
This research reveals VLM outputs are highly sensitive to context, presenting a material challenge for consistent and reliable enterprise deployments.
Do what
Your model risk and validation frameworks must specifically account for context-dependent variability in VLM outputs before broader adoption of multimodal AI.