{"slug": "grasp-reinforcing-language-model-anonymizers-with-group-relative-policy", "title": "GRASP: Reinforcing Language Model Anonymizers with Group Relative Policy Optimization", "summary": "Researchers introduced GRASP (Group-Relative Anonymization via Self-refinement Policy-optimization), a method that uses Group Relative Policy Optimization to train a small on-device language model for adversarial anonymization, improving the privacy-utility trade-off over DPO-distilled baselines. Trained on Llama-3.1-8B, GRASP matches or outperforms frontier-model-driven anonymization from Gemini 2.5 Flash and Claude while removing more private information, and runs at roughly 1% of the GPT-4o teacher's cost.", "body_md": "arXiv:2608.06526v1 Announce Type: new\nAbstract: Large language models can infer sensitive personal attributes, such as age, location, and occupation, from ordinary text, turning everyday writing into a privacy risk. Adversarial anonymization defends against this by rewriting a text with a capable language model that also plays the attacker, but it needs a powerful model at inference time and thus sends private text to a third party, the very exposure anonymization should prevent. Recent work distills this behavior into a small on-device model using supervised fine-tuning and direct preference optimization (DPO), but DPO only imitates the teacher's offline choices and never directly optimizes the privacy--utility objective we care about. We introduce \\textbf{GRASP} (\\textbf{G}roup-\\textbf{R}elative \\textbf{A}nonymization via \\textbf{S}elf-refinement \\textbf{P}olicy-optimization), which reinforces the local anonymizer online with Group Relative Policy Optimization. A single small model acts as anonymizer, adversary, and utility judge, trained against a self-generated reward that hides attributes while preserving meaning, with a design that guards against reward hacking. Trained on Llama-3.1-8B, \\ours{} improves the privacy--utility trade-off over the DPO-distilled baseline, consistently across three independent LLM judges. Against adversarial anonymization driven by frontier models such as Gemini~2.5~Flash and Claude, it achieves a comparable or better overall trade-off while removing substantially more private information, and it runs entirely on-device at roughly $1\\%$ of the GPT-4o teacher's cost.", "url": "https://wpnews.pro/news/grasp-reinforcing-language-model-anonymizers-with-group-relative-policy", "canonical_source": "https://arxiv.org/abs/2608.06526", "published_at": "2026-08-10 04:00:00+00:00", "updated_at": "2026-08-10 04:11:08.417627+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "ai-safety"], "entities": ["GRASP", "Llama-3.1-8B", "Gemini 2.5 Flash", "Claude", "GPT-4o"], "alternates": {"html": "https://wpnews.pro/news/grasp-reinforcing-language-model-anonymizers-with-group-relative-policy", "markdown": "https://wpnews.pro/news/grasp-reinforcing-language-model-anonymizers-with-group-relative-policy.md", "text": "https://wpnews.pro/news/grasp-reinforcing-language-model-anonymizers-with-group-relative-policy.txt", "jsonld": "https://wpnews.pro/news/grasp-reinforcing-language-model-anonymizers-with-group-relative-policy.jsonld"}}