Learning Style, Forgetting Semantics
An academic paper published on arXiv that provides a theoretical explanation, via exact decomposition of a linear-softmax policy, for why supervised fine-tuning (SFT) is more prone to catastrophic forgetting than reinforcement fine-tuning (RFT). The central finding is that style drift, rather than semantic errors, is the primary source of forgetting.
Timeline (last 90 days)
The arXiv paper "Learning Style, Forgetting Semantics" explains the phenomenon of SFT being more prone to forgetting than RFT through theoretical modeling.
The paper focuses on classification task scenarios, where different tokens within the same semantic category express the same semantics but exhibit different styles.
The authors verified the theoretical predictions through simulation experiments on task sequences.
All Facts (3)
The paper focuses on classification task scenarios, where different tokens within the same semantic category express the same semantics but exhibit different styles.
50%UnverifiedThe authors verified the theoretical predictions through simulation experiments on task sequences.
50%UnverifiedThe arXiv paper "Learning Style, Forgetting Semantics" explains the phenomenon of SFT being more prone to forgetting than RFT through theoretical modeling.
50%