New CHaRS method steers language models to reduce toxicity

Researchers have developed a new method for steering language model activations, called Clustered Optimal-Transport Steering (CHaRS). This technique uses optimal transport to pair clusters of classes and moves hidden states based on their proximity to these clusters, offering an alternative to the common practice of adding a single vector. Experiments on Gemma-2-2B and Granite-3-1-2B models showed that CHaRS, particularly a variant that routes tokens to their nearest clusters, significantly reduced toxicity compared to the standard method, even under INT8 and NF4 quantization. AI

IMPACT Introduces a novel technique for mitigating toxicity in language models, potentially improving their safety and reliability.

RANK_REASON Academic paper detailing a new method for language model activation steering. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• CHaRS
• Gemma-2-2B
• Granite-3-1-2B
• INT8
• Mean-AcT
• Toxic-Bert

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →