We introduce i-CIR—an instance‐level composed image retrieval benchmark with rigorously curated hard negatives—and BASIC, a training‐free VLM‐based method that centers and projects image embeddings.