Anchor: Mitigating Artifact Drift in Agent Benchmark Generation 事件

Anchor: Mitigating Artifact Drift in Agent Benchmark Generation arXiv:2605.26321v1 Announce Type: new Abstract: AI agents are beginning to complete valuable, long-horizon business operations tasks, but training and evaluation environments for enterprise work still struggle to balance realism, verifiability, and scale. Environment and task creation frequently suffers from a failure mode we call artifact drift: when instructions, environments, oracles, and verifiers are created by loosely coupled