BenchmarkCertified Human-Robot Intelligence
CEHRI
A domain examination benchmark comprising 60 questions across three task categories: factual questions, arithmetic operations, and implicit-goal reasoning. It is used to evaluate a language model's error-correction capability and its retention of general performance after correction.
Timeline (last 90 days)
Sep 16
On the CEHRI 60-question domain exam, CRN v2 corrected 53.3% of the base model's errors, and 43.3% on paraphrase variants.
Unverified50%