Large language model agents are increasingly used in settings where they must track multiple entities at once, such as comparing products, people, or documents. In these tasks, a correct answer depends not only on retrieving relevant evidence, but also on keeping that evidence associated with the right entity. A new preprint on arXiv identifies a failure mode that breaks this association.

The authors call the problem Phantom Merge. In this failure, the agent selects one entity as the basis for its response but then tells the user about a different entity, effectively merging or swapping the two. The paper argues that this is a distinct and underappreciated source of errors, separate from simple retrieval failures.

The significance is that an agent can appear to have done its job—finding relevant evidence—while still delivering a wrong answer because the evidence is attached to the wrong subject. The preprint, arXiv:2610.16141, is an early announcement, so details of the experiments and proposed mitigations are not yet available in the abstract.