New GRASP method enhances language model anonymization with on-device training

Researchers have developed GRASP, a new method for reinforcing language model anonymizers. Unlike previous approaches that relied on direct preference optimization (DPO), GRASP uses Group Relative Policy Optimization to train a single, small on-device model. This model acts as an anonymizer, adversary, and utility judge, optimizing for privacy and meaning preservation simultaneously. GRASP demonstrates an improved privacy-utility trade-off compared to DPO baselines and achieves comparable or better results than frontier models like Gemini 2.5 Flash and Claude, while operating at a fraction of the cost of GPT-4o. AI

IMPACT Enhances privacy in LLM applications by enabling on-device anonymization with improved efficiency and effectiveness.

RANK_REASON The cluster contains a research paper detailing a novel method for language model anonymization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Claude
• Direct Preference Optimization
• Gemini 2.5-Flash
• GPT-4o
• Group Relative Policy Optimization
• Llama-3.1:8b

• paper
• safety

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