Li, C., He, H., Gao, Y., Li, M., Ye, J., Yang, Q., & Ye, P. (2026). Rethinking Token Reweighting for SFT: Suppress, Reverse, and Extrapolate Learned Features. https://omanscience.com/en/articles/rethinking-token-reweighting-for-sft-suppress-reverse-and-extrapolate-learned-features