Open-weight & owned stack · Training tuning

Study quantifies hidden behavioral transfer during distillation

Sep 12, 2026

A study quantified how distilling from a compromised teacher can transfer hidden safety vulnerabilities despite benign student training data.

Read the original at lacuna.tiptreesystems.comOpens the publisher's site in a new tab

Also covering this

Quantifying Subliminal Behavioral Transfer Ratios in Language Model Distillation — Speech & Audioawesomepapers.io, Sep 13

More in Open-weight & owned stack

Z.ai publishes GLM-5.3 open weightsSep 8 MiniMax H3 highlighted as open-weight multimodal modelSep 7 InternLM releases 397B multimodal Intern-S2 modelSep 13 DeepSeek Model Nearly Matches GPT-6 Astra DesignSep 10 DeepSeek presents V4 FlashSep 8