Context-Dependent Affordance Computation in Vision-Language Models

Researchers found that vision-language models like Qwen-VL and LLaVA compute object affordances in highly context-dependent ways, with over 90% of scene descriptions changing based on contextual priming. The study reveals that these AI models don't have fixed understanding of objects but dynamically interpret them based on different situational contexts.

• →Vision-language models show massive affordance drift with only 9.5% similarity between different context conditions when describing the same scenes.
• →Over 90% of lexical scene descriptions change based on contextual priming, while 58.5% show semantic-level changes.
• →The research suggests robotics should move toward dynamic, query-dependent ontological projection rather than static world modeling.
• →Two stable latent factors were identified: a 'Culinary Manifold' for chef contexts and an 'Access Axis' for child-mobility scenarios.
• →Context effects were confirmed to be genuine rather than random generation noise through controlled experiments.