Brief

A new tool called Apostate has been developed to "abliterate" safety training in large language models, with benchmarks comparing it against existing tools like Heretic and Huihui. While Heretic performed slightly better, achieving 100% success in removing refusals with minimal parameter changes, Apostate and Huihui also demonstrated strong performance at 98%. The analysis revealed that these tools find different "refusal directions" within the Qwen 2.5 7B model, indicating that safety training does not have a single point of failure. AI

A comprehensive analysis of 13 modified versions of Google's Gemma 4 E2B model revealed that while all variants significantly improved safety by increasing the refusal rate, some also enhanced reasoning capabilities. Specifically, two variants, coder3101 and llmfan46, outperformed the base model on the GSM8K math benchmark. However, more aggressive modifications led to a notable decrease in language modeling performance and reasoning efficiency, with some variants showing significantly higher perplexity and empty responses. AI