Summary
- Researchers propose an operational measure called Natural-Language-Rooted (NLS) depth to quantify unverbalized serial cognition in AI models.
- NLS depth measures the length of the longest path through computation without an interpretable bottleneck, evaluated via architecture and training configuration.
- The measure correlates with layer count across open-weight models and remains insensitive to ordinary post-training methods like prompting or RL.
- Latent-reasoning architectures could significantly increase NLS depth compared to standard transformers, impacting AI model monitorability.
Why it matters
As models evolve toward latent-reasoning architectures, tracking metrics like NLS depth helps developers assess and maintain AI monitorability and alignment oversight.