What's this about?
Inference-time scaling improves reasoning, but usually through repeated generation or an external verifier — both expensive. CritICL instead learns from a *small* model's failure modes to improve a strong model's inference-time behaviour.
What is "weak-to-strong generalisation"?
The counterintuitive finding that a weaker model can be used to improve a stronger one. The intuition: a small model's mistakes are informative. They map out where a problem is genuinely hard versus where it merely looks hard, and that map can steer the strong model's effort. It is the same reason a teacher learns more from students' wrong answers than from their right ones.
Why does this matter to you?
- It avoids the two standard costs of inference-time scaling — many samples, or a separate verifier model.
- Weak-to-strong is also a live safety question: it is the shape of humans overseeing systems more capable than themselves.
- Small-model failure modes are cheap to collect, which makes this practical rather than merely elegant.