Most deployed personalization systems assume models will faithfully infer only what evidence supports. This paper shows the opposite is common: LLMs routinely invent user attributes beyond the available evidence, and models' own self-reports can mislead model-level comparisons.
Key Findings
- MirageBench setup: 150 balanced personas (stereotypical / counter-stereotypical / neutral), 6 personalization tasks spanning an "imagination gradient", and a four-way faithfulness taxonomy judged by an independent verifier.
- Scale of evaluation: 12 models from 7 families, 143,616 judged claims. Models over-infer at 35%–49% per model (cross-model mean 41.6%; claim-weighted 41.8%).
- Self-Monitoring Inversion: models that self-report lower over-inference tend to be measured as fabricating more (model-level rank correlation rho = -0.60, p = 0.044). Within-model self-audit still provides moderate ranking ability (AUROC 0.58–0.83).
- Task and temporal dynamics: OI varies by task (27%–59%) and inferred attributes accumulate roughly linearly across turns with little revision in a pilot multi-turn study.
Who it's for and tradeoffs
Great fit if you evaluate or deploy LLM personalization, design memory-augmented agents, or need an empirical benchmark for profile faithfulness: MirageBench provides a large, judge-validated evaluation and a leaderboard to compare external verification metrics. Look elsewhere if you need operational deployment guidance or mitigation code—this work diagnoses behaviors and supplies evaluation tooling rather than turnkey fixes. Be aware the findings reflect the tested models and tasks; results guide trustworthy personalization by prioritizing external verification over opaque self-reports.
