AI & Tech··1 sources verified
Researchers Propose NLS Depth Metric to Track Hidden Reasoning in AI Models
Summarised by Relevant News AI · Read time: 3 min
A new arXiv paper defines NLS depth, a proxy for how much serial reasoning a model can perform without expressing it in human-readable chain-of-thought. The measure is set by architecture and training recipe, so it can be assessed before training and is unaffected by prompting, distillation, or RL. The authors find NLS depth tracks layer count in open-weight models and could rise sharply with latent-reasoning architectures, and they urge developers to report it alongside direct chain-of-thought monitorability measures.
Why it matters: If future architectures reason in latent space, chain-of-thought oversight may stop working, so developers need a pre-training metric to decide how much of a model's reasoning they can actually monitor.
All sources