PrivAgent is an efficient AI agent architecture for real-time privacy risk monitoring in sensitive environments. Developed in collaboration with the NHS (UK National Health Service) Sandbox and submitted to ACL, it solves the critical challenge of ensuring AI agents comply with privacy regulations like HIPAA and GDPR at every action — in real time, not as an afterthought.
Key Features
Rule Condensation — An LLM compiler compresses verbose legal texts (HIPAA's 144 articles, GDPR's 99 articles) into ~30-token typed checkers, achieving 10–50x compression.
Two-Pass Short-Circuit Detection — Phase 1 concurrently checks forbidden rules (immediate reject on violation); Phase 2 checks warning rules (all must pass to proceed).
SFT + GRPO Training — Specialized training on Qwen3-8B raises lenient accuracy from 47.5% to 88.8%, with an 8B model outperforming the 70B baseline across precision, speed, and efficiency.
Sub-Second Compliance — Prefix-cache-aware scheduling with sglang RadixAttention achieves 0.48–0.51s per request (vs. 10–14s for the 70B baseline).
Results
- Accuracy improved by +4.0–9.5% (Medical) and +23.9–24.8% (CARES-18K) over 70B baseline
- 21–27x faster on the same hardware
- 46–54% token reduction
- Latency: 0.48–0.51s/req (baseline: 10–14s)
Collaborators
NHS (UK National Health Service) Sandbox
PrivAgent 是面向隐私敏感环境的高效 AI Agent 实时隐私风险监控架构。与英国国家医疗服务体系（NHS）Sandbox 合作，投稿 ACL。解决 AI Agent 在医疗等隐私敏感场景下，如何在每一次动作执行时实时、高效、可解释地完成 HIPAA/GDPR 等隐私法规合规检查。
核心功能
规则压缩 — LLM 编译器将冗长的法律条文（HIPAA 144 条/GDPR 99 条）压缩为 ~30 token 的类型化 Checker，实现 10–50× 压缩。
两阶段短路检测 — Phase 1 并发检查 Forbidden List（DENY/OBEY 规则），任一违规立即拒绝；Phase 2 检查 Warning List（CHECK 规则），全部通过才放行。
SFT + GRPO 专项训练 — 在 Qwen3-8B 上训练，宽松准确率从 47.5% 提升至 88.8%。8B 模型全面超越 70B 基线。
亚秒级合规检查 — Prefix-cache-aware 调度 + sglang RadixAttention，0.48–0.51s/req（70B 基线 10–14s）。
成果
- 精度对比 70B 基线 +4.0–9.5%（Medical）、+23.9–24.8%（CARES-18K）
- 同硬件加速 21–27×
- Token 节省 46–54%
- 延迟 0.48–0.51s/req（基线 10–14s）
合作方
NHS（英国国家医疗服务体系）Sandbox