[KongchangAI]
ConceptEIG / Expected Information Gain

Expected Information Gain (EIG)

A core metric in Bayesian experimental design that quantifies how much uncertainty about a hidden state a candidate query is expected to reduce. It is defined as the expected entropy reduction of the belief distribution before and after observation.

Timeline (last 90 days)

Nov 3

In matched teacher comparison experiments, EIG supervision consistently outperformed the random query strategy.

Unverified50%
Nov 3

Expected Information Gain (EIG) is the core metric of Bayesian experimental design, used to quantify how much uncertainty about the hidden state a candidate question can reduce in expectation.

Unverified50%

All Facts (2)

Source Articles