VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions — ThinkLLM