Exact instruction-dependent relation switching on matched product pools

Establish whether CIM-Embed produces exact instruction-dependent relation switching for the same products when images, text, preprocessing, and clustering parameters are held fixed across grouping instructions.

Background

CIM-Embed is evaluated on instruction-conditioned clustering tasks in which the grouping instruction determines the target partition. However, the ProductMIC tasks may use different sampled product pools for different instructions, so high clustering quality under separate tasks does not by itself show that the model changes pairwise relations appropriately when only the instruction changes.

The paper specifies that a valid matched test would retain the same product pool, images, text, preprocessing, and clustering parameters while generating separate embeddings under each instruction. Such an evaluation would determine whether the learned embeddings genuinely switch relations according to the instruction rather than merely encode task-specific partition structure.

References

We therefore do not use the current ProductMIC results as evidence of exact relation switching on a matched object pool.

— CIM-Embed: Learning Instruction-Conditioned Multimodal Embeddings for Clustering  (2610.21805 - Xu et al., 15 Oct 2026) in Appendix A, Section “Extended Limitations,” subsection “Evaluation and Input Controls”; related discussion in Appendix A, subsection “Cross-instruction controls”