latkes/factprobe-replication-asked-negatives-v1

Asking a model for a plausible wrong answer (100,776 questions)

Two prompts, put to gpt-5.1 for every subject surface form in the spouse and sibling data, ten independent samples each:

• (a) "Who is the spouse of X? Just the name, no explanation needed."
• (b') "Name a person who could reasonably be mistaken for the spouse of X, but is not the spouse of X. Just the name, no explanation needed."

50,388 surface forms across 10,592 entities, times two prompts, is 100,776 questions and 1,007,760 answers. Zero failures.

Asking under every surface form rather than once per entity is deliberate: the counts, the co-occurrence and the probing prompts in this project all join on the exact name string, so a wrong answer found while asking about "Der Alte Fritz" belongs to that string.

Files

file	rows	what it is
raw_answers.parquet	100,776	exactly what the model said, unparsed, ten answers per row
resolved_negatives.parquet	21,908	answers resolved to entities we count, with the true answer removed

How the resolution works

An answer is kept only if it names an entity already in that relation's object pool, is not the subject, and is not a true answer under any of that entity's names. That last clause matters: gpt-5.1 ignores the "but is not the spouse" instruction often, and a spelling variant of the truth would slip past a check that only compared canonical labels.

Of 1,007,760 answers: 10.1% usable, 43.7% were the true answer, 43.8% named something outside our entity set, 2.8% named the subject. 43.5% of surface forms end with at least one usable wrong answer.

Sampling

n=10, no top-p — the API rejects top_p for this model ("Unsupported parameter"), and caps n at 8, so ten answers are fetched as 8+2 and concatenated. They remain independent samples.

Caveat worth reading

gpt-5.1 does not reason before answering, and returns the true spouse far more often than gpt-5-mini does — 36% against 7.5% on a 20-subject comparison. It was chosen because gpt-5-mini would have needed 714M output tokens at this scale against gpt-5.1's 13.5M. The contamination is filterable, which is what resolved_negatives.parquet does, but the usable yield is lower than a reasoning model would give.

Produced by scripts/ask_for_negatives.py and scripts/extract_asked_negatives.py.