What's this about?
FlyBy lets a small model think first, diagnose whether it lacks execution skill or knowledge, and selectively query a stronger model only when useful.
What exactly is a knowledge bottleneck?
It is a failure caused by missing facts rather than insufficient reflection. Thinking longer cannot recover information that was never in the model.
Why does this matter?
21 HF upvotes. The paper reports better pass@8 than a 14B comparison model at 2.7× lower serving cost, suggesting routing decisions can matter as much as model size.