In plain words
An agent with long-term memory may rely on an old fact when taking an action. The same fact might be fine for a routine task but risky when a mistake would be costly.
The paper introduces AMV-Guard, a gate that checks the facts behind a proposed action and decides whether to proceed, check, refuse, or ask the agent to replan. In simulated tool-use episodes, it used less verification effort than a version that did not adapt its checks, while both recorded no harmful executions among cases the gate could evaluate. These results apply to the simulated tests.
The research
When should an agent trust a stored fact enough to act on it? A memory’s age or confidence alone cannot answer that question: the same fact may be adequate for a routine action and need checking when a mistake would have greater consequences. The paper asks how to decide when to check, proceed or abstain.
AMV-Guard checks the stored claims behind each proposed action when a check could change whether an agent proceeds, abstains or replans.
AMV-Guard checks the claims behind proposed actions when a check could change whether an agent proceeds, abstains or replans.
In ToolSandbox, a simulated tool-use test where agents act on tasks and an evaluator scores outcomes, it was compared with AMV-Static, which did not adapt its checks to new observations.
Across three observation seeds, AMV-Guard lowered mean verification cost per episode from 1.843 to 1.219. Both systems had zero harmful executions among cases the gate could evaluate.
Both also executed every ground-truth-safe original candidate the gate could evaluate during the initial Gate episode, possibly after verification and before replanning. AMV-Guard’s end-to-end raw harm was slightly higher. The comparison supports lower cost under this test setup, not better results on every measure.
These tests use simulated environments and cover only the consequences their evaluators score. The AppWorld result transfers the decision core, not AMV-Guard’s full premise-extraction and checking process.
Analysis
Deep
- Original paper authors
- Yiqi Wang, Yifei Wang, Jiaqi Zhang, Zhangkai Wu, Yiqun Duan
- Submitted to arXiv
- License
- arXiv non-exclusive