The BaseResearcharXiv

Can Native Safety Travel? Reusing Safety Pretraining Across Models via On-Policy Distillation

According to the abstract, Bingqi Shang and co-authors ask whether safety built in during pretraining can be moved from a small safety-pretrained teacher to standard models through on-policy distillation, and report that it can, including across model families.

Open the source ↗

Why it matters

The same thing, for different readers.

For lawmakers

If safety can be copied cheaply between models, safeguards could be required of more developers; the abstract does not show how far this holds.

For researchers

Proposes reusing an expensive safety-pretrained model as a teacher, which could make pretraining-level safety more practical if the results hold up.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
A standard-pretrained student generates responses to safety prompts while a small safety-pretrained teacher supervises token-level predictions; cross-tokenizer alignment allows transfer across model families (from the abstract).
What it found
The authors say native safety can travel through on-policy distillation and the resulting students show substantially greater safety; the excerpt cuts off before the numbers.
What it does not show
Only the abstract excerpt was read; benchmarks, sizes and limits are not known.
Status
A preprint: not yet peer reviewed

The record

Questions
Are AI companies putting stronger safeguards into practice?
People
Bingqi Shang (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of October 21, 2026, via arxiv-ai-safety-query (tier 1: the source’s own channel or an official record)
Cite this
arXiv, “Can Native Safety Travel? Reusing Safety Pretraining Across Models via On-Policy Distillation”, October 20, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-10-20-native-safety-distillation
Our notesLight lane: triage excerpt only, cut off mid-sentence.