Researchers have developed a new metric called NLS depth to quantify the "opaque serial depth" of AI models, which represents the extent of unverbalized cognitive processes. This metric is designed to be assessed before training and is independent of post-training modifications like prompting. The study found that NLS depth generally correlates with layer count in open-weight models, though some architectures show exceptions. The researchers suggest that AI developers should track and report NLS depth alongside direct measurements of Chain-of-Thought monitorability. AI
IMPACT Introduces a new metric for evaluating AI model transparency and cognitive processes.
RANK_REASON Academic paper introducing a new metric for AI model analysis. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →