Concept评分幻觉

GRADE-HALLUCINATION

A specific failure mode in LLM-as-Judge evaluation settings, in which the judge model exhibits erroneous sensitivity to surface-level features such as numeric formatting, column name style, and result ordering. This leads to a systematic bias in which correct outputs are incorrectly classified as unfaithful.

Timeline (last 90 days)

Sep 28

The researchers named the core mechanism of judge misjudgment GRADE-HALLUCINATION, referring to the judge developing erroneous sensitivity to surface features and thus misjudging correct outputs

Unverified50%

All Facts (1)

Unverified

The researchers named the core mechanism of judge misjudgment GRADE-HALLUCINATION, referring to the judge developing erroneous sensitivity to surface features and thus misjudging correct outputs

50%

Source Articles

Auditing LLM Judges: Evaluation Failures and Fixes in Text-to-SQL PipelinesSep 28