Knowledge distillation is a standard technique for compressing large models by training a smaller student model on the outputs of a larger teacher. It is valued for transferring capabilities, but a new arXiv paper (2610.26311) highlights a growing concern: distillation can also transfer unintended teacher properties, including backdoors.

The paper, titled 'Selective Transfer through Sequence Distillation: From Benchmark Knowledge to Backdoor Behaviour,' sets out to study whether different hidden teacher characteristics—specifically backdoor behaviours—are passed on through sequence distillation. The abstract notes that recent work has already shown such unintended transfer, but the full findings are not available in the announcement.

Because the source is a single arXiv abstract with the results section truncated, this digest can only report the paper's motivation and scope. No experimental outcomes or comparisons are described in the provided text.