Published on June 16, 2026, Next-Latent Prediction (NextLat) is presented as a self-supervised learning method that trains transformers to predict their next latent state.
In a publication dated June 16, 2026, Next-Latent Prediction (NextLat) is described as a self-supervised learning method for training transformers to predict their next latent state. The publication claims the approach enables compact world models and inference up to 3.3 times faster with self-speculative decoding.
The method may interest engineers exploring alternative training objectives and faster inference for transformers. To verify the result, consult the original publication and check how it defines the method, measures speed, and compares results; details of those tests are not included in the available summary.