LineageEval

This research page investigates whether political censorship inherent in Chinese frontier AI models transfers to smaller models distilled from them. The study uses DeepSeek V4 Flash as a teacher and an American model (GPT-OSS-120B) as the student, focusing on financial reasoning as a representative use case. The key finding is that while the student gains performance in the target domain, it does not inherit the teacher's censorship behaviors: across 152 matched prompt pairs, the distilled model showed no statistically significant difference in censorship from the untouched base model, whereas the teacher scored 45.45 points more censored on China-sensitive questions than on structurally identical controls. The page also demonstrates that self-distillation (training on the model's own corrected continuations) yields similar performance gains, and releases the evaluation tool LineageEval for reproducibility.