Researchers at Nous Research have successfully retrained Alibaba's Qwen3.8-27B model, creating Arete-27B. This process significantly reduced the model's pro-Chinese political bias from 22% to under 1% through supervised fine-tuning on 53 million tokens. Notably, the model's general capabilities remained largely unaffected by this bias reduction. AI
IMPACT Demonstrates that political bias in AI models can be a superficial layer, potentially removable without compromising core functionality.
RANK_REASON Research paper detailing a method to reduce bias in an AI model. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
