Research map / Reinforcement learning / Visual and representation RL
Learning Efficient Guardrails for Compliance
Xiaofei Wen, Wenjie Jacky Mo, Yanan Xie, Peng Qi, Muhao Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Qwen3-4B-Instruct on PolicyGuardBench | Qwen3-4B-Instruct · PolicyGuardBench | accuracy: 0.6897 -> 0.9014 | Improved |
| fine-tune Qwen3-4B-Instruct on PolicyGuardBench | Qwen3-4B-Instruct · PolicyGuardBench | F1: 0.5348 -> 0.8759 | Improved |
| use safety guardrail models (LlamaGuard, ShieldGemma) instead of policy-specific guardrail | policy compliance detection · PolicyGuardBench | accuracy: 0.9014 -> 0.4256 | Hurt |
| scale up model from 8B to 70B in Llama family | Llama · PolicyGuardBench | accuracy: 0.6647 -> 0.9054 | Improved |
| lengthen prefix from N=1 to N=5 (ablation) | PolicyGuard-4B · PolicyGuardBench prefix | accuracy: 0.9101 -> 0.8190 | Hurt |
| lengthen prefix from N=1 to N=5 (ablation) | Llama-3.2-3B-Instruct · PolicyGuardBench prefix | accuracy: 0.9086 -> 0.5693 | Hurt |
| hold out one domain (out-of-domain test) instead of in-domain (ablation) | PolicyGuard-4B · PolicyGuardBench LODO | accuracy: 0.9328 -> 0.9083 | Hurt |
| deploy PolicyGuard-4B as online interceptor | web agent · Closed-loop web-agent simulation | Completion-Under-Policy: 10.20% -> 16.33% | Improved |
Closest papers by meaning
- ShieldAgent: Shielding Agents via Verifiable Safety Policy Reasoning (ICML 2025 · similarity 0.90)
- LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails (NeurIPS 2026 · similarity 0.89)
- LiSA: Lifelong Safety Adaptation via Conservative Policy Induction (NeurIPS 2026 · similarity 0.89)
- Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.88)
- SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations (ICLR 2025 · similarity 0.87)
- GuardAgent: Safeguard LLM Agents via Knowledge-Enabled Reasoning (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
