Knowing When Thinking Is Not Enough: Teaching Small Reasoning Models to Reason Beyond Their Parametric Knowledge
The paper’s claim is that small reasoning models need to learn when to stop self-refining and ask a stronger model for missing knowledge.
The authors separate failures into execution bottlenecks, where reflection can recover a reachable path, and knowledge bottlenecks, where outside information is needed. Their FlyBy framework trains 4B and 8B models to reason first, diagnose what is unresolved, and selectively query stronger models. On 1,158 hard problems across six benchmarks, FlyBy-4B reports 45.96% pass@8, ahead of Qwen3-14B at lower serving cost. FlyBy-8B reaches 51.81% pass@8. HF Daily Papers' note
The authors separate failures into execution bottlenecks, where reflection can recover a reachable path, and knowledge bottlenecks, where outside information is needed. Their FlyBy framework trains 4B and 8B models to reason first, diagnose what is unresolved, and selectively query stronger models. On 1,158 hard problems across six benchmarks, FlyBy-4B reports 45.96% pass@8, ahead of Qwen3-14B at lower serving cost. FlyBy-8B reaches 51.81% pass@8. HF Daily Papers' note
score 5