▼Training Objectives Move Beyond Next-Token Prediction

Score history

Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.

Now 38 · -1 since Sep 13 · ranged 38 to 39

Showing the last few days. Unlock full score history.

Why the conviction moved

• Sep 13
  Strengthened +5

  LeVJEPA replaces the usual self-supervised heuristics in latent-space video prediction with a single regularizer that provably rules out representation collapse. Removing the main theoretical objection to joint-embedding predictive objectives makes the training objective, rather than data or parameter count, the lever producing the gain.

• Sep 11
  Strengthened

  NCP-ArchPreview scales latent-space concept-level prediction to 8.9 billion parameters, beating OLMo-3 on reasoning benchmarks including a 5.99-point GSM8K improvement, the largest published run of this objective class to date.

Source trail

• Supporting · September 13, 2026

  A Collapse-Free Video Encoder Matches V-JEPA 2 With Up to 20.8 Times Less Pretraining Compute

  LeVJEPA replaces the usual self-supervised heuristics in latent-space video prediction with a single regularizer that provably rules out representation collapse. Removing the main theoretical objection to joint-embedding predictive objectives makes the training objective, rather than data or parameter count, the lever producing the gain.

  AI with Papers (Telegram)

• Supporting · September 11, 2026

  A Latent-Space Language Model Scales to 8.9 Billion Parameters and Beats OLMo-3 on Reasoning

  NCP-ArchPreview scales latent-space concept-level prediction to 8.9 billion parameters, beating OLMo-3 on reasoning benchmarks including a 5.99-point GSM8K improvement, the largest published run of this objective class to date.

  arXiv cs.CL