Redwood Defines 'NLS Depth' for AI Monitoring; Distinguishes Kimi-K3's High Depth from DeepSeek's Overestimation

Redwood Research defined the 'NLS depth' metric based on Natural Language-rooted nodes (NL-rooted nodes) to measure 'opaque serial depth' within AI models. This metric is based on the depth of computational paths that do not pass through NL-rooted bottlenecks. According to the source, the opaque serial depth of standard Transformers using CoT (Chain-of-Thought) tends to be proportional to the number of layers. Kimi-K3, which uses Gated DeltaNet, was analyzed to have a higher NLS depth per layer compared to the standard Transformers used for comparison. DeepSeek-V4-Pro, which does not use DeltaNet, also showed high values, but this is distinguished as a case where depth was overestimated due to Sinkhorn-Knopp operations. The researchers pointed out that this operation, while serial, has low FLOPs and does not use learned parameters, identifying the high metric values as a flaw in the current NLS depth definition. Additionally, they revealed that most post-training methods, such as prompt engineering or general CoT training, keep nodes in an NL-rooted state and therefore do not increase NLS depth.