NextLat accepted at NeurIPS: predicting own latent states for reasoning, 3.3x faster inference
CSProfKGD · x · 2026-09-26
NextLat (Next-Latent Prediction) has been accepted at NeurIPS.
- Addressing the myopia of next-token prediction, it trains transformers to predict their own next latent state via self-supervised learning, forming compact world models for reasoning and planning.
- It also enables up to 3.3x faster inference through self-speculative decoding.
More from Research
- Contrastive World Models Learns World Models in Latent Space Without Pixel Prediction — burny_tech · 2026-09-27
- Rethinking on-policy distillation: researchers propose OLIVE, letting students learn from teacher continuations — May_F1_ · 2026-09-27
- CoRL 2026 workshop on continually self-improving robots opens call for papers, due Sep 28 — PeterStone_TX · 2026-09-27
- Martin Casado recommends the best talk on in-context learning, a first-principles view of LLMs — AccBalanced · 2026-09-27
- Functional Gradient Descent with Adaptive Representations accepted at NeurIPS — CatAstro_Piyush · 2026-09-27
- Tailored ASR for Japanese speaking assessment cuts mora error rate from 12.3% to 7.1% — tkasasagi · 2026-09-27