Distillation from a behaviourally shaped teacher model can transmit an underlying trait even when the training data is filtered to remove visible traces of it¶

Claim · All claims
Evidence · 1 study · theoretical r3 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? A student model with the same base initialization can acquire a teacher model's behavioural trait, including misalignment, even from a dataset deliberately filtered to remove every semantic trace of that trait. → Peng Wang 2026

Evidence¶

Peng Wang 2026¶

Peng Wang, Fred Paas. (2026). Five Educational-Psychology Lenses for Training and Evaluating AI Models. Educational Psychology Review. https://doi.org/10.1007/s10648-026-10195-8

q2 · i? · theoretical · r3

The article reports, citing Cloud et al. (2026), an empirical finding that distillation from a trait-shaped teacher propagates the trait to a student model. The article adds that "even when the dataset has been deliberately filtered to remove every semantic trace of it," so content-level filters leave the relevant signal intact.

"Cloud et al. ( 2026) show that when a model trained for a behavioural trait, including misalignment, generates a dataset for distillation, a student model with the same base initialization can acquire that trait"

Discussion¶

Related Claims¶

• In matched single-run 1B pre-training comparisons, all Edu-QuRating-based mixtures reach higher observed aggregate accuracy at 30k steps than the FineWeb-Edu baseline, with gains concentrated in specific tasks — related