Unverified50% confidenceBenchmarkExact time
Research by the University of Bonn and Fraunhofer IAIS shows that the same model scored 0.73 on a reasoning benchmark when using the raw text, dropped to 0.52 after personal details were redacted, and rose back to 0.70 after personal information was replaced with consistent pseudonyms.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn the author's internal document corpus, the median Top-1 score for questions with true answers using the ada-002 model is approximately 0.84, while between random unrelated chunks it is approximately 0.72.69% similarUnverifiedThe study found that the faithfulness score for counterfactual inputs was only 0.05 points lower than for factual inputs (on a 1–5 scale), a difference that is almost negligible.68% similarUnverifiedIn manually proofread samples, the character error rate was only 0.68% when the text was clearly legible, but approximately 45% of sampled passages were completely uncorrectable due to source quality issues.68% similarUnverifiedAt the actual repetition levels of ordinary text, five models ranging from 1B to 13B parameters carry only a faint trace of what they have seen, with a rank correlation coefficient of approximately -0.08.67% similarUnverifiedIn 10%–25% of cases, reasoning about the hypothetical grader pulled the agent's work away from the user's original specification.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026392API
curl https://kongchang.com/api/v1/knowledge/claims/1026392MCP
get_claim(id=1026392)