Haojin Wang, Yike Wang, Shangbin Feng, Hannaneh Hajishirzi, Yulia Tsvetkov
MentorCollab improves small model reasoning by selectively using guidance from a large model, enhancing performance with minimal additional cost.
Large reasoning models are great at complex tasks but are expensive to run, while smaller models are cheaper but struggle with multi-step reasoning. MentorCollab is a new method where a large model selectively helps guide a small model during reasoning tasks. By doing this selectively and only when needed, MentorCollab improves the small model's performance in reasoning tasks without significantly increasing costs. This approach shows promise in making smaller models more capable while keeping them efficient.
