BenchmarkEvaluation-Invariant Concept Stability
EICS
A metric for evaluating the representational stability of vision-language models. It tests whether a model's internal representations consistently encode a target concept through single-concept intervention, revealing the model's true representational quality.
Timeline (last 90 days)
Oct 12
EICS (Evaluation-Invariant Concept Stability) is a metric for probing representations that tests whether a model's representations can consistently encode a target concept through controlled single-concept interventions
Unverified50%