Paper
18 min
Thinking longer, not bigger
A 70B model given four minutes to think beats a model ten times its size answering at once. Here is why, and where it stops working.
M. Castell, J. Okafor, T. Lindqvist
The usual way to make a model smarter is to make it bigger. We show that for hard problems, giving a model more time to search, test and revise is a cheaper lever.
On our hardest maths and science set, accuracy rises from 41% at one second to 92% at thirty minutes. A model ten times larger, answering immediately, reaches 63%.
What we found
Gains come mostly from dropping bad branches early
Self-checking matters more than longer chains
Returns flatten on tasks with no way to verify an answer
The code and evaluation harness are open. Everything needed to reproduce the curves in this paper is in the repository.
Keep reading
More from the lab.
System card
32 min
Kerrfield One system card
How Kerrfield One was trained, what 11,400 hours of red-teaming found, and every mitigation we shipped before launch.
Paper
18 min
Thinking longer, not bigger
A 70B model given four minutes to think beats a model ten times its size answering at once. Here is why, and where it stops working.
Paper
14 min
Calibration you can check
When Kerrfield One says it is 80% sure, it is right about 80% of the time. How we trained for honest confidence, and how to verify it yourself.