Self-Privileged Fine-Tuning (SPFT)

Yangrong Hu, Xueyu Zhou, Yue Wu, Jian Huang.

Under review at ICLR 2027.

Overview

Privileged-information-guided token weighting uses a frozen, reference-conditioned model without external teachers or on-policy rollouts, improving mean mathematical reasoning accuracy from 36.41% to 47.80% over standard SFT on Qwen3-8B-Base.

All publications & manuscripts →