TLDR Tech / AI
A Mirror Test For LLMs
A new 'Mirror Test' evaluates LLM self-awareness by requiring models to identify their own outputs without direct hints. Anthropic's Opus 4.6 demonstrated significant self-recognition due to unique token outputs, surpassing OpenAI's GPT models, which failed to recognize their own generations. While some LLMs showed signs of attempted self-marking, none consistently exhibited self-awareness through effective message passing.
Read the full story