BLAST RADIUS

Open-weight & owned stack · Training tuning

Study Uses Zeroth-Order Optimization To Harden LLM Safety

Sep 11, 2026

A paper proposed zeroth-order optimization to improve language-model safety alignment against noise and quantization while preserving utility.

Read the original at awesomepapers.ioOpens the publisher's site in a new tab

More in Open-weight & owned stack

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