A new paper on arXiv examines a core assumption in the design of universal multimodal embeddings (UMEs): that adding reasoning will make retrieval better. The authors note that reasoning-enhanced UMEs do improve heterogeneous retrieval, but they also observe that plausible rationales do not necessarily lead to discriminative rankings. In other words, a model can produce a convincing explanation while still failing to rank relevant items sharply above irrelevant ones.
The paper sets out to study this gap by comparing discriminative (DISC) performance, though the abstract cuts off before giving details of the method or results. Because the full text is not yet available, the precise findings remain unclear. What is clear is that the authors are pushing back against the assumption that reasoning is always useful in this setting.
The preprint is available on arXiv under the identifier 2609.29560.