LLMs still leak real-item knowledge in low-knowledge persona simulations
Original title: Can LLMs Ignore What They Know? Measuring Knowledge Leakage in Persona-Based User Simulation
Across five API endpoints from two model families, explicit ignorance prompts still left a favourable-item real-twin rating gap.
Researchers define knowledge leakage as bias when a simulated persona uses knowledge the real user could not have, and measure it with real items paired to fabricated twins. In rating tests across three knowledge-handling arms, personas did not fully suppress real-item existence and familiarity signals, even after an in-context statement that the persona had never heard of the item. Pooled unfavourable items sat near zero, but domain groups reversed sign. Estimates depend on twin construction, and human low-knowledge baselines are still needed.
Unconfirmed: All estimates are conditional on twin construction and screening; human low-knowledge baselines are still required.
- Favourable real-twin rating separation persisted under explicit-ignorance prompts.
- Pooled unfavourable contrast was +0.077, with interval −0.057 to +0.204.
- Hazard ratings returned −0.359; company or organisation ratings returned +0.438.
- Five API endpoints from two model families were tested.
From the original article
1 min read at arxiv.orgWhen persona-conditioned responses are read as respondent-level simulations, each response should reflect the described persona's knowledge, and the respondent written to know nothing must act as if the model's knowledge were not there. We test whether it can, and find persistent favourable-item real-twin rating separation under the tested low-knowledge and explicit-ignorance prompts.
By Sangmin Lee · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.