Add new paper:

[wyzh0912]

Title

Understanding and Mitigating Gender Bias in LLMs via Interpretable Neuron Editing

Published Date

2025-01-24

Source

arXiv

Head Name

Gender Head

Summary

• Innovation: The paper introduces an interpretable neuron editing method to mitigate gender bias in LLMs by targeting specific biased neurons while preserving the model’s original capabilities. This approach combines logit-based and causal-based strategies to selectively edit neurons responsible for gender bias.

• Tasks: The study involves the creation of the CommonWords dataset to evaluate gender bias across five LLMs and employs a detailed analysis of neuron circuits to identify "gender neurons" and "general neurons" that contribute to gender bias, followed by experiments to test the proposed neuron editing method.

• Significant Result: The proposed method effectively reduces gender bias in LLMs while maintaining their original capabilities, outperforming existing fine-tuning and neuron editing approaches across several benchmarks, as demonstrated by improvements in ICAT scores and reductions in entropy differences.