Devin Fusion: Frontier Performance at 35% Lower Cost

2 pointsposted 11 hours ago
by fanyangxyz33

1 Comments

fanyangxyz33

11 hours ago

I'm very curious about how they trained the lightweight classifier to decide model switching. Is it supervised? Did they use LLM as a teacher? It also seems that the featurization isn't trivial. Like how you build a simple but still meaningful representation of the task(s).