A new training approach teaches small AI models exactly when to ask a bigger model for help, and what to ask, instead of handing off whole chunks of a problem.
The method reframes small-model/large-model (SLM-LLM) teamwork as an information-acquisition problem under a fixed API budget, rather than the usual approach of splitting reasoning work between two models. The small model stays in charge of the actual reasoning and only queries a larger, black-box LLM advisor when it hits a specific gap, sending a targeted question instead of forwarding the whole task. To make that work, the researchers built a three-stage reinforcement-learning framework that trains the small model to decide when to call the advisor, how to phrase a useful query, and how to fold the answer back into its own reasoning. On math and coding benchmarks, the approach beat existing collaboration methods on the cost-versus-performance tradeoff, and in some cases matched or beat an oracle baseline that routes every problem to the ideal model from the start.
Most SLM-LLM setups treat the big model like a subcontractor: pass it a chunk of the problem, get an answer back, repeat. That is slow and expensive when every call costs money. Training the small model to ask sharp, specific questions instead keeps most of the big model's reasoning boost while using far fewer tokens, and the technique reportedly carries over to other advisor model families without retraining.
It is still a lab result on math and coding benchmarks, not a product. And teaching a model to know when it needs help is a hard skill - plenty of people never master it either.