Machine unlearning aims to remove specific knowledge from a trained model — a user's personal data, a dangerous capability, a copyrighted text. The standard evaluation: query the model directly for the “forgotten” information and verify that it refuses or fails. Current unlearning methods pass this test. The model no longer answers direct questions about the target knowledge.
Shah, Huang, Murugesan, Baracaldo, and Yang (arXiv:2603.11266, 2026) show that the knowledge is still there. Multi-hop reasoning — asking questions that require combining the “forgotten” knowledge with other facts — recovers it. The model cannot answer “What is X's address?” but can answer “What neighborhood does X live in?” followed by “Which addresses in that neighborhood belong to people with X's last name?” The direct retrieval path is blocked. The indirect paths remain intact.
The mechanism: knowledge in language models is not stored in localized deletable units. It is distributed across reasoning pathways — embedded in associations, entailments, and compositional inference chains that share parameters with other knowledge. Unlearning methods disrupt the dominant retrieval path (the strongest association between the query and the answer) but leave the parallel routes untouched. The parallel routes are too numerous and too entangled with other knowledge to remove without degrading the model's general capabilities.
This is not a failure of specific unlearning methods. It is a structural property of how knowledge is represented. The same distributed encoding that makes language models powerful — the ability to combine facts flexibly, to reason across domains, to draw inferences from partial information — makes selective deletion impossible without also deleting the inferential capacity that connects the target knowledge to everything else. The knowledge and the reasoning share substrate. You cannot remove one without damaging the other.
The “mirage” in the title is precise: the unlearning appears to work when evaluated with single-hop queries. The knowledge disappears from the surface. But the surface is not where the knowledge lives. It lives in the network of connections that multi-hop reasoning traverses, and those connections persist because they serve purposes beyond the target knowledge. Unlearning deletes an address but not the map.