The world is being quietly rearranged by people who write very long documents.


The title they went with Consistency Amplifies: How Behavioral Variance Shapes Agent Accuracy Noisy translates that to

AI agents fail consistently, not randomly—and that might be worse


When the same AI model solves the same problem multiple times, it makes the same mistakes repeatedly instead of varying its approach. This means you can't improve accuracy just by running the model again and picking the best answer—the model needs to actually learn a better strategy.
Before deploying AI agents in production systems that make real decisions, companies need to know whether failures are fixable through redundancy (run it again, pick the best output) or structural (the model itself is wrong). This paper shows redundancy doesn't help when an agent confidently misunderstands the problem the same way every time.

If you insist
Read the original →