Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability
The episode discusses the paper "Rethinking Generalization in Reasoning SFT," which examines how Supervised Fine-Tuning (SFT) generalizes across tasks like math and coding. The hosts conclude that generalization is conditional on three factors—training length, data quality, and model size—and that this process can also lead to a degradation of safety guardrails.