Fair-GPTQ: Bias-Aware Quantization for Large Language Models

Name: Fair-GPTQ: Bias-Aware Quantization for Large Language Models
Uploaded: 2026-10-11T09:40:29.321100+00:00
Duration: 698 s
Description: Fair-GPTQ is a new method to reduce bias in large language models during compression (quantization). It modifies the standard quantization objective by adding a regularization term that penalizes large errors when comparing inputs that differ only by stereotypical versus anti-stereotypical attributes. This aims to produce less biased outputs while maintaining model efficiency.

Fair-GPTQ is a new method to reduce bias in large language models during compression (quantization). It modifies the standard quantization objective by adding a regularization term that penalizes large errors when comparing inputs that differ only by stereotypical versus anti-stereotypical attributes. This aims to produce less biased outputs while maintaining model efficiency.

Open the full episode