TLDR Tech / Practical
AI models link code analysis to morality experts
Researchers studying mixture-of-experts language models discovered that when assessing malicious code, these models activate a path associated with morality rather than purely technical code assessment. Tests on various open models revealed that questions about malice heavily engage this morality path, which is positioned closely to morality-related queries in the model's routing. Further experiments indicated this path serves as a lookup mechanism rather than storing the judgment itself.
Read the full story