AdaGuard: An Adaptive Guard Model with User-defined Policies
AdaGuard guard models judge agent behavior under user-defined policies; the 4B model hits 89.3% accuracy.
AdaGuard is a family of 0.6B, 4B, and 8B guard models that score agent trajectories against policies supplied at inference time rather than a fixed risk taxonomy. Training uses AdaptiveSafety, with 10,939 training examples and 1,000 test examples covering policies of 1–100 rules, plus policy and behavioral counterfactuals. SafePO then refines violation detection with structured rewards that separately weight explanations and verdicts. The 4B model reaches 89.30% binary accuracy on AdaptiveSafety and 71.82% on DynaBench.
- AdaptiveSafety has 10,939 training and 1,000 test examples with 1–100 rules.
- Counterfactuals and rule-order augmentations teach policy-dependent judgments.
- SafePO balances explanation and verdict rewards with a separate value model.
- AdaGuard ships 0.6B, 4B, and 8B models that read policies at inference time.
- The 4B model scores 89.30% on AdaptiveSafety and 71.82% on DynaBench.
- AI models
- AdaGuard-0.6BAdaGuard-4BAdaGuard-8B
From the original article
1 min read at huggingface.co
Guard models support the safe deployment of language model agents, but fixed risk taxonomies limit their ability to accommodate requirements that vary across applications and tasks. Under user-defined policies, detecting violations requires interpreting both the applicable rules and the agent's behavior, since identical actions can receive different judgments under different policies.
By Yunhao Feng, Yifan Ding, Yuxiang Xie · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.