A model is 'relatively biased' when its responses deviate from the consensus of a baseline LLM set, and this deviation can be scored by embedding distances or LLM judges plus equivalence tests.
Amazon bedrock guardrails
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Relative Bias: A Comparative Framework for Quantifying Bias in LLMs
A model is 'relatively biased' when its responses deviate from the consensus of a baseline LLM set, and this deviation can be scored by embedding distances or LLM judges plus equivalence tests.