The problem. In multi-turn personalized localization, a system holds stored visual evidence — for instance, images of a person or place — and must decide whether that evidence applies to the current request. The authors of a new arXiv preprint argue that this relevance is not a fixed property of the evidence. It can change as the interaction state changes, even when the stored evidence itself remains identical.

The contribution. To make this problem tractable, the paper introduces SCVIB, a controlled diagnostic benchmark designed to study state-dependent source relevance. The benchmark is the central contribution described in the abstract; details of its construction and results are not covered in the available source text.

Why it matters. By separating the stored evidence from its state-dependent relevance, SCVIB targets a gap in evaluating multi-turn personalized systems. The work is available as arXiv:2608.14148 (v2, replacement).