RealFeedApp beta

GAMA-Bench tested 1,298 scenarios for LLMs

Researchers developed the GAMA-Bench benchmark to evaluate whether large language models apply consistent standards to negative behavior performed by men and women. The test includes 1,298 scenarios covering both intimate and public social conflicts.

Field
training language models
What they did
Researchers created a dataset of 1,298 conflict scenarios where identical actions are performed by actors of different genders to test if models judge them consistently. They developed a method to measure how AI models distribute punishment, empathy, and blame.
Why it matters
This will help understand why models often show harshness toward men and allow testing their ability to remain fair in various situations.
#llm#gender bias#gama-bench#moral framing#evaluation

Read the original →