Scaling test-time computation has become a popular route to stronger language-model reasoning, especially for small reasoning models (sRMs) that are inexpensive to serve. But a new paper asks whether more thinking is always the right move. By intervening at intermediate reasoning states across two model families and multiple scales, the authors find that self-refinement mostly reshuffles probability mass onto solutions already reachable from the current state rather than opening new ones.
Those interventions expose two distinct failure regimes. In execution bottlenecks, the correct path is internally reachable and reflection can recover it. In knowledge bottlenecks, progress requires external information that the model simply does not possess.
To address the second regime, the researchers introduce FlyBy, a selective querying framework. Models are trained to reason first, diagnose what remains unresolved, and — at a knowledge bottleneck — query stronger models whose parametric knowledge extends beyond their own. Supervised fine-tuning bootstraps a multi-depth query action, while cost-aware reinforcement learning calibrates whether to query, what to ask, and how much to spend.
On 1,158 hard problems spanning six benchmarks, FlyBy-4B reaches 45.96% pass@8, surpassing Qwen3-14B (41.64%) at 2.7 times lower serving cost, and also edges out Qwen3-8B in pass@1 (16.85% vs. 15.31%). Scaling to FlyBy-8B lifts pass@8 further to 51.81%.
The work frames selective querying as a way to integrate external information into a small model's own reasoning without paying the cost of indiscriminate querying.