AI Latent Reasoning: Sep-Latent Token Compression

By applying a progressive Attention-Decay mask that selectively starves intermediate reasoning tokens of attention during training, Sep-Latent forces the transformer to compress multi-step logical states into structural separator anchors, and aligns these compressed states with an MLP-projected continuous latent stream via all-layer Euclidean self-distillation. Paired with a dynamic LM-head exit gate that autonomously terminates silent computation based on query complexity, Sep-Latent keeps latent vectors grounded on the vocabulary manifold, cuts training GPU memory by over 50%, and achieves