Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• ICE-Guard is a framework that uses intervention consistency testing to detect demographic, authority, and framing biases in LLMs.method 95%
• Authority bias and framing bias in LLMs are substantially higher than demographic bias.result 95%
• Bias concentration varies significantly across different domains, with finance exhibiting higher authority bias than criminal justice.result 95%
• Structured decomposition, involving LLM feature extraction and a deterministic rubric, can significantly reduce LLM flip rates.method 95%
• An ICE-guided detect-diagnose-mitigate-verify loop can achieve substantial cumulative bias reduction through iterative prompt patching.method 95%
• Real COMPAS recidivism data shows higher flip rates than pooled synthetic rates, suggesting the benchmark is a conservative estimate of real-world bias.result 95%
• The study evaluated 11 LLMs from 8 different families.technical 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card