AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Learning Efficient Guardrails for Compliance

Xiaofei Wen, Wenjie Jacky Mo, Yanan Xie, Peng Qi, Muhao Chen

ICML 2026 · Reinforcement learning › Visual and representation RL · approach: RL for combinatorial routing · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
fine-tune Qwen3-4B-Instruct on PolicyGuardBenchQwen3-4B-Instruct · PolicyGuardBenchaccuracy: 0.6897 -> 0.9014Improved
fine-tune Qwen3-4B-Instruct on PolicyGuardBenchQwen3-4B-Instruct · PolicyGuardBenchF1: 0.5348 -> 0.8759Improved
use safety guardrail models (LlamaGuard, ShieldGemma) instead of policy-specific guardrailpolicy compliance detection · PolicyGuardBenchaccuracy: 0.9014 -> 0.4256Hurt
scale up model from 8B to 70B in Llama familyLlama · PolicyGuardBenchaccuracy: 0.6647 -> 0.9054Improved
lengthen prefix from N=1 to N=5 (ablation)PolicyGuard-4B · PolicyGuardBench prefixaccuracy: 0.9101 -> 0.8190Hurt
lengthen prefix from N=1 to N=5 (ablation)Llama-3.2-3B-Instruct · PolicyGuardBench prefixaccuracy: 0.9086 -> 0.5693Hurt
hold out one domain (out-of-domain test) instead of in-domain (ablation)PolicyGuard-4B · PolicyGuardBench LODOaccuracy: 0.9328 -> 0.9083Hurt
deploy PolicyGuard-4B as online interceptorweb agent · Closed-loop web-agent simulationCompletion-Under-Policy: 10.20% -> 16.33%Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy