KnowBias: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• KnowBias mitigates social bias in LLMs by strengthening neurons encoding bias-knowledge rather than suppressing neurons associated with biased behavior.method 95%
• KnowBias achieves state-of-the-art debiasing performance across multiple benchmarks and LLMs.result 90%
• KnowBias requires only a handful of simple yes/no questions and no retraining, making it highly data efficient.technical 90%
• KnowBias preserves general model capabilities while achieving debiasing, minimizing utility degradation.result 85%
• KnowBias generalizes across different bias types and demographic groups.technical 85%
• KnowBias relies on the assumption that bias knowledge is consistently encoded in neurons across different models, which may not be universally true.limitation 90%
• Existing debiasing methods that adopt a suppressive paradigm are often brittle, weakly generalizable, data-inefficient, and prone to degrading general capability.technical 80%
• KnowBias could replace or augment existing debiasing technologies that focus on neuron-level suppression, offering a more robust and efficient solution.market 75%

Competitive landscape

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

Startup potential card