LLM◆ AI-generated · Sourced
LLMs Know Your Project, Not You: NameRank Measures Parametric Recognition
TL;DR
An arXiv paper uses NameRank (a 0-1 recognition score) to measure what a model recalls about a person or tool from its own weights before any retrieval; the finding: models recognize named, indexable artifacts, not credentials or titles.
Bottom line
What a frontier model recalls about a person or tool from its weights—before any retrieval—often shapes the first description a human sees. NameRank makes that a measurable problem.
How it's measured
- 4,685 entities across 54 cohorts, each probed with one open-ended question across 36 models.
- An independent judge gives a binary verdict: did the model state a specific, non-guessable fact about this exact entity? Hallucination, context echo, and guesses earn nothing.
- Synthetic-null entities hold the floor near zero.
One key finding
Recognition is paid to named, indexable artifacts—not to credentials or titles.
- Olympic-style honors sit below a working-researcher baseline, because a medal ships with no named artifact.
- But it inverts at the marquee tier: Nobel, Turing, and Fields laureates saturate the panel.
- For independent creators, the tool out-ranks its maker; the credential that propagates is a named method or awarded paper.
For teams doing personal/organizational GEO (generative engine optimization) who want to be cited by AI search, the guidance is clear: ship named, indexable artifacts.
Sources (compliance trail)
https://arxiv.org/abs/2607.12520