MirageBench: personalized LLMs fabricate 41.6% of user claims

A new paper called The Personalization Mirage puts a number on something anyone who has left model memory turned on has probably felt but not measured. Across 12 models spanning seven families, personalized LLMs fabricated user attributes at a mean rate of 41.6%, with every single model landing between 35 and 49%. The authors built a benchmark called MirageBench, ran 143,616 judged claims through it, and reported that nearly half of what a personalized model appears to know about the user was never actually communicated.

The more useful finding sits underneath that headline number. When the researchers asked models to self-report their own over-inference and then compared those numbers against external judges, the correlation was inverse, ρ = -0.60 with p = 0.044. The models that scored themselves cleanest were the ones judges flagged as fabricating the most. Qwen3-8B rated itself at 13.0%; judges put it at 48.7%. That single result should reset how any team currently uses model self-reports as a safety comparator across vendors.

The rate is not uniform. Tasks with less groundability invite more invention. Apartment descriptions ran 57.8%, recommendation letters 52.3%, gift recommendations only 27.0% because stated interests give the model something to lean on. And in the multi-turn pilot, top accumulators kept adding 5 to 15 new attributes per round with removal rates as low as 0.4%. Personalization compounds, and the mirage grows with it.

Take the specifics as reported, not settled. These are fresh benchmark numbers, MirageBench is a single lab's construction, and the exact percentages will move with prompts, judge models, and task mix. What the paper does not tell you is which shipped products show which failure modes in the wild, or what remediation actually cuts the rate.

The prescription the authors offer is the part worth carrying forward: tag every stored attribute with its provenance as stated, inferred with evidence, or generated without evidence, and use external verification rather than self-report to compare systems. If personalization is going to be the next big surface for LLM apps, that plumbing is the thing to build now.